# AI Sandbox Hunt: every fact > Ratings, prices and limits for agent sandboxes and other compute for AI agents. 103 providers, facts last checked 2026-11-15. Every price comes from the vendor's own page; the source URLs are listed. Scores use the published method at https://aisandboxhunt.com/methodology. Free to reuse with a link to the provider page. Short index: https://aisandboxhunt.com/llms.txt. JSON: https://aisandboxhunt.com/api/v1/providers.json. # AI agent sandboxes Isolated environments, also called AI code sandboxes, that an AI agent starts through an API or SDK to run code, usually for seconds to hours, then throws away. Ranking: https://aisandboxhunt.com/code-sandboxes ## Isorun Isorun runs each sandbox as a microVM with its own kernel, booted from any OCI image and driven from a TypeScript SDK, an MCP server or the API. Running sandboxes bill per second at $0.025 per vCPU-hour and $0.015 per GiB-hour, so the default 1 vCPU / 1 GiB size is $0.04 an hour; hibernated sandboxes are not billed for runtime. Hibernate keeps memory, processes and files, and optional persistent disks keep a workspace across sandboxes for $0.02 per GB-month of data stored. The service is in public beta. - Page: https://aisandboxhunt.com/isorun - Company: Isorun - Website: https://isorun.ai - Categories: AI agent sandboxes, Dev boxes for AI agents - Score: 78/100, #1 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $19.36 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.025 - Price per GiB-hour of memory: $0.015 - Storage per GiB-month: $0.02 - Billing increment: per second, no minimums - Billed while: Wall-clock from create to destroy (or the auto-destroy timer) while running; hibernated sandboxes are not billed for runtime. Persistent disks bill for data stored. - Plans: Pay as you go $0.00/month (No plan fee; per-second usage. Concurrency limits are not published) - Free tier: $50 of free credit for new accounts, no card required (about 1,250 hours of the default 1 vCPU / 1 GiB sandbox). - Isolation: KVM microVM per sandbox with its own guest kernel - Pause and resume: yes - Snapshots: yes - Custom images: yes - Network controls: yes - Max vCPU: 8 - Max memory GiB: 16 - SDKs: TypeScript, MCP server, REST API - Regions: US or EU; each organization is scoped to one region chosen at sign-up. - Open source: no - Self-hostable: no - Best for: Short code-execution runs from any Docker image with per-second billing and no minimum; Agent workspaces that hibernate between sessions and resume with processes and memory intact; Locking down egress with allow-lists and keeping API keys out of the sandbox - Watch out for: Public beta; Sandboxes are destroyed after 5 idle minutes by default unless the timeout is changed or the sandbox is hibernated; Only a TypeScript SDK is offered; other languages use the API or MCP server; Retention of hibernated sandboxes is not documented - Benchmark (ComputeSDK benchmarks, 2026-11-13): Burst start time (median of 100 at once) 96 ms, rank 1 of 27 - Benchmark (ComputeSDK benchmarks, 2026-11-13): Repo clone, install and typecheck (one run) 36034 ms, rank 2 of 25 - Sources: https://docs.isorun.ai/getting-started/pricing https://isorun.ai https://docs.isorun.ai/sandboxes/persistent-storage https://docs.isorun.ai/sandboxes/lifecycle https://docs.isorun.ai/getting-started/authentication - Facts checked: 2026-11-14 ## Northflank Northflank sandboxes are Northflank services created through its API or JavaScript/Python SDKs; on Northflank Cloud, CPU sandboxes run in Cloud Hypervisor microVMs and GPU sandboxes under gVisor. Compute is billed per second against the chosen resource plan at $0.01667 per vCPU-hour and $0.00833 per GB-hour of memory, with persistent volumes and egress charged separately. Sandboxes have no time limit, can be paused to stop compute billing, and can also run on BYOC clusters in your own cloud account. - Page: https://aisandboxhunt.com/northflank - Company: Northflank - Website: https://northflank.com/product/sandboxes - Categories: AI agent sandboxes, GPU sandboxes - Score: 78/100, #2 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $11.73 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.01667 - Price per GiB-hour of memory: $0.00833 - GPU prices per hour: NVIDIA L4 24 GB $0.80, NVIDIA A100 40 GB $1.42, NVIDIA A100 80 GB $1.76, NVIDIA H100 80 GB (8x only) $2.74, NVIDIA RTX PRO 6000 96 GB $3.00, NVIDIA H200 141 GB (8x only) $3.14, NVIDIA B200 180 GB (8x only) $5.87 - Storage per GiB-month: $0.1166 - Billing increment: per second - Billed while: Wall-clock while the sandbox has running instances, for the CPU and memory of its resource plan; paused sandboxes (scaled to zero) pay only for attached volumes. - Plans: Sandbox (free) $0.00/month (No card needed; up to $20/month of resources on Northflank Cloud across 2 services, 2 jobs and 1 database; 1 BYOC cluster up to 12 vCPU and 48 GB; limited vertical scaling, no horizontal scaling or autoscaling); Pay as you go $0.00/month (Usage-based with no seat fees; unlimited users, deployments and workloads; GPUs, autoscaling, RBAC and the Platform SDK) - Free tier: Free Sandbox plan: up to $20/month of resources on Northflank Cloud (2 services, 2 jobs, 1 database), no card needed. - Isolation: Cloud Hypervisor microVM (CPU sandboxes); gVisor (GPU sandboxes) - Pause and resume: yes - GPU: yes - Custom images: yes - Network controls: yes - Max vCPU: 288 - Max memory GiB: 2048 - SDKs: JavaScript, Python, CLI, REST API - Regions: Northflank Cloud regions chosen per project (GPU sandboxes need a GPU-enabled region such as asia-southeast), or BYOC on AWS, Google Cloud, Azure, Oracle, CoreWeave, Nebius and others. - Open source: no - Self-hostable: yes - Best for: Long-running or always-on agent sandboxes, since there is no time limit; Large or GPU sandboxes (up to 288 vCPU, L4 to B200 GPUs); Teams that want the same sandbox API on Northflank Cloud or a BYOC cluster in their own cloud - Watch out for: Billed for the full CPU and memory of the selected plan while running, even when idle; pause the sandbox to stop compute charges; Persistent volumes bill per GiB-month even while paused; egress costs $0.06/GB; Free plan is capped at $20/month of resources, 2 services, 2 jobs and 1 database, with no horizontal scaling; A project's region cannot be changed later, and GPU sandboxes need a GPU-enabled region - Sources: https://northflank.com/pricing https://northflank.com/product/sandboxes https://northflank.com/docs/v1/application/sandboxes/sandboxes-on-northflank-cloud - Facts checked: 2026-11-14 ## Koyeb Sandboxes Koyeb Sandboxes are ephemeral environments created from Docker images with Koyeb's Python and JavaScript SDKs, running on the same microVM instances as Koyeb services. Each sandbox is billed per second at the price of its instance type, from $0.0036 per hour for 0.25 vCPU / 256 MB up to GPU instances such as an H100 at $2.50 per hour. Sandboxes scale to zero after an idle timeout and can be auto-deleted after up to 24 hours, and filesystem or full snapshots can seed new sandboxes. Sandboxes are in public preview and need a Starter (pay-as-you-go), Pro or Scale plan. - Page: https://aisandboxhunt.com/koyeb-sandboxes - Company: Koyeb - Website: https://www.koyeb.com/docs/sandboxes - Categories: AI agent sandboxes, GPU sandboxes - Score: 76/100, #3 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $10.14 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Fixed-size price per hour: $0.0144 (small: 1 vCPU / 1 GB / 10 GB SSD) - GPU prices per hour: NVIDIA RTX 4000 SFF Ada 20 GB $0.50, NVIDIA L4 24 GB $0.70, NVIDIA RTX A6000 48 GB $0.75, NVIDIA L40S 48 GB $1.20, NVIDIA A100 80 GB $1.60, NVIDIA RTX PRO 6000 96 GB $2.20, NVIDIA H100 80 GB $2.50, NVIDIA H200 141 GB $3.00, NVIDIA B200 180 GB $5.50, 2x NVIDIA H100 80 GB $5.00, 8x NVIDIA H100 80 GB $20.00 - Billing increment: per second - Billed while: Per second at the instance type's price while the sandbox's instance runs; after the idle timeout (default 300 seconds) it scales to zero and compute stops. - Plans: Starter $0.00/month (Pay as you go with a valid payment method; no base fee); Pro $29.00/month ($10 of compute included each month. 10 users, 100 services, CPU concurrency 100, GPU concurrency 20, 1 TB/month bandwidth); Scale $299.00/month ($100 of compute included each month. 50 users, 1,000 services, CPU concurrency 1,000, GPU concurrency 40, AWS regions, 99.9% uptime SLA) - Free tier: No free sandbox usage is stated. Koyeb's one free instance is for a single web service and is not offered for sandboxes. - Isolation: MicroVM per instance on bare-metal servers - Pause and resume: yes - Snapshots: yes - GPU: yes - Custom images: yes - Network controls: yes - Max session hours: 24 - Max vCPU: 40 - Max memory GiB: 128 - SDKs: Python, JavaScript/TypeScript - Regions: Koyeb regions in the US, EU and Asia (default 'na'); GPU instances only in specific regions. - Open source: no - Self-hostable: no - Best for: Sandboxes that need a GPU, from an L4 to multi-GPU H100 and H200 machines; Teams already deploying services on Koyeb; Bursty sandboxes that should scale to zero between requests - Watch out for: Still in public preview; No free sandbox usage; a payment method or paid plan is required; Auto-deletion is capped at 24 hours after creation; Storage pricing for snapshots is not published on the pages checked - Sources: https://www.koyeb.com/docs/reference/instances https://www.koyeb.com/docs/sandboxes https://www.koyeb.com/docs/reference/organizations https://www.koyeb.com/docs/sandboxes/sandbox-lifecycle https://github.com/koyeb/koyeb-python-sdk/blob/main/docs/sandbox.md https://www.koyeb.com/blog/koyeb-sandboxes-javascript-sdk-now-available - Facts checked: 2026-11-13 ## Lizard Sandboxes Lizard Sandboxes are Firecracker microVMs with their own kernel, created from TypeScript or Python SDKs, the lizard CLI, the dashboard or an HTTP API in US East or EU West. Each size has a flat hourly price that covers CPU, RAM and local disk, from $0.00405 for 0.5 vCPU / 1 GB to $0.1296 for 16 vCPU / 32 GB, billed per second including idle time. Pausing stops compute charges and keeps the files, but not memory or running processes; a paused sandbox costs $0.14 per GB-month of used disk. Running sandboxes can be forked with memory included. Usage comes from a Pro plan that costs $19 a month and includes $19 of usage. - Page: https://aisandboxhunt.com/lizard - Company: Lizard - Website: https://lizard.build - Categories: AI agent sandboxes - Score: 76/100, #4 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $19.00 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Fixed-size price per hour: $0.0081 (Mini: 1 vCPU / 2 GB RAM / 10 GB disk) - Storage per GiB-month: $0.14 - Billing increment: per second - Billed while: Wall-clock while the sandbox runs, idle time included. Pausing stops compute charges; a paused sandbox and each snapshot cost $0.14 per GB-month of used disk. Persistent Volumes bill separately. - Plans: Pro $19.00/month ($19 of usage a month across apps, databases and sandboxes, then pay as you go. Up to 40 vCPU of sandboxes running at once, unlimited starts); Pro 5x $99.00/month ($99 of usage a month. Up to 220 vCPU of sandboxes at once, unlimited starts); Pro 20x $499.00/month ($499 of usage a month. Up to 1,200 vCPU of sandboxes at once, unlimited starts) - Free tier: 7-day Pro trial with $5 of usage; a card is required but nothing is charged during the 7 days. The trial runs at most 1 Small, 2 Mini and 2 Nano sandboxes at a time, in US regions only, and deletes each after at most 1 hour. - Isolation: Firecracker microVM per sandbox with its own kernel - Pause and resume: yes - Snapshots: yes - GPU: no - Custom images: no - Max session hours: 596 - Max vCPU: 16 - Max memory GiB: 32 - SDKs: JavaScript/TypeScript, Python, CLI, REST API - Regions: US East (Virginia, us-east-1) and EU West (Limburg), chosen per sandbox. The trial runs only in US regions. - Open source: no - Self-hostable: no - Best for: Large fleets of small sandboxes where a low flat hourly price matters more than idle-time billing; Agents that fork a running sandbox with memory and processes included; Teams that also want to host apps and databases on the same plan - Watch out for: Running sandboxes bill for idle time; pausing stops compute but keeps billing for used disk, and memory is not kept; Pro caps sandboxes at 40 vCPU running at once; No custom image upload; templates come from Lizard's catalog; The trial needs a card and limits sandboxes to small sizes in the US - Sources: https://lizard.build/pricing https://lizard.build/docs/platform/limits/ https://lizard.build/docs/sandboxes/ https://lizard.build - Facts checked: 2026-11-15 ## Tensorlake Sandboxes Tensorlake runs CPU sandboxes in Firecracker microVMs, created from Python, TypeScript, the tl CLI or the API. Ephemeral sandboxes run until terminated or timed out; named sandboxes can be suspended and resumed with memory, processes and filesystem kept, and can be used as SSH dev environments. CPU is billed on active execution time with a floor of 20% of allocated vCPU, while RAM and root disk are billed on allocation while the sandbox runs. Plans range from a free tier of one small sandbox to prepaid credits and a $250/month Pro plan with lower rates. - Page: https://aisandboxhunt.com/tensorlake - Company: Tensorlake - Website: https://tensorlake.ai - Categories: AI agent sandboxes, Dev boxes for AI agents, GPU sandboxes - Score: 76/100, #5 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $17.95 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.07 (active CPU only) - Price per GiB-hour of memory: $0.015 - GPU prices per hour: NVIDIA H100 (reserved capacity, separate invoice) $3.95 - Storage per GiB-month: $0.07 - Billing increment: per second, billed in whole compute units ($0.01 each), pooled per organization - Billed while: CPU: active execution time, with a minimum of 20% of allocated vCPU (full allocation when active metering is unavailable). RAM and root disk: allocated amount while running. Suspended sandboxes pay snapshot storage only. - Plans: Free $0.00/month (1 sandbox at a time, 1 vCPU, 1 GB RAM, 10 GB disk, sessions up to 2 hours. Sessions are not metered; no paid usage is possible on this plan); Usage Credits $0.00/month (Prepaid packs of $5, $10 or $20 (500-2,000 compute units); no postpaid overage, sandboxes stop when the balance hits zero unless auto-refill is on. Up to 100 concurrent sandboxes, 4 vCPU, 16 GB RAM, 100 GB disk, sessions up to 24 hours); Pro $250.00/month ($250/month including 25,000 compute units ($250) of usage at Pro rates, then $0.01 per unit. 1,000 concurrent sandboxes, up to 16 vCPU, 64 GB RAM, 100 GB disk, no session time limit. HIPAA with a BAA) - Free tier: Free plan at $0 with no credit card: 1 sandbox at a time with 1 vCPU, 1 GB RAM and 10 GB disk, sessions up to 2 hours, unmetered. - Isolation: Firecracker microVM per sandbox (GPU sandboxes use gVisor) - Pause and resume: yes - Snapshots: yes - GPU: yes - Custom images: yes - Network controls: yes - Max vCPU: 16 - Max memory GiB: 64 - SDKs: Python, TypeScript, CLI, REST API - Regions: Region list not published; the docs state EU data residency is supported. Enterprise can run in your own AWS, GCP or Azure account. - Open source: no - Self-hostable: yes - Best for: Agent tool calls and harnesses that wait on model calls, since only active CPU time is billed; Named sandboxes used as SSH or IDE dev environments that suspend when idle and resume with processes intact; Evals and RL rollouts that fork a warmed sandbox into many copies - Watch out for: RAM and root disk are billed on allocation for the whole run, and idle CPU still bills 20% of the allocated vCPU; Pay-as-you-go sandboxes are limited to 4 vCPU, 16 GB and 24-hour sessions; larger sizes and unlimited sessions need the $250/month Pro plan; GPU sandboxes need a request; the only published GPU price is reserved H100 capacity; Region list is not published - Benchmark (ComputeSDK benchmarks, 2026-11-13): Burst start time (median of 100 at once) 991 ms, rank 16 of 27 - Benchmark (ComputeSDK benchmarks, 2026-11-13): Repo clone, install and typecheck (one run) 72954 ms, rank 21 of 25 - Sources: https://tensorlake.ai/pricing https://docs.tensorlake.ai/platform/billing https://docs.tensorlake.ai/sandboxes/introduction https://docs.tensorlake.ai/sandboxes/lifecycle https://docs.tensorlake.ai/sandboxes/gpu https://docs.tensorlake.ai/sandboxes/networking https://tensorlake.ai - Facts checked: 2026-11-14 ## Vercel Sandbox Vercel Sandbox runs each session in a Firecracker microVM, created from the JavaScript/TypeScript or Python SDK or the sandbox CLI. CPU is billed only for time the code actively uses it, while memory is billed for the full provisioned amount for as long as the session runs (2 GB per vCPU). Sandboxes are persistent by default: stopping one snapshots its filesystem (billed as snapshot storage) and the next call resumes it in a new session of up to 24 hours on Pro. Hobby includes a monthly usage allowance at no cost; Pro usage draws from the plan's $20 monthly credit. - Page: https://aisandboxhunt.com/vercel-sandbox - Company: Vercel - Website: https://vercel.com/sandbox - Categories: AI agent sandboxes - Score: 76/100, #6 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $28.44 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.128 (active CPU only) - Price per GiB-hour of memory: $0.0212 - Storage per GiB-month: $0.08 - Billing increment: Active CPU by CPU time used; provisioned memory in 1-minute minimum increments - Billed while: CPU: only while code actively uses it (I/O and model-call waits are free). Memory: full provisioned amount for the whole session, including idle time. Nothing while stopped, except snapshot storage. - Plans: Hobby $0.00/month (Monthly allowance (not a dollar credit): 5 Active CPU hours, 420 GB-hours of memory, 5,000 sandbox creations, 20 GB data transfer; 15 GB snapshot and 15 GB drive storage (lifetime). 10 concurrent sandboxes, 45-minute sessions, max 4 vCPU / 8 GB. Sandbox creation pauses when the allowance is used up); Pro $20.00/month ($20/month platform fee (1 deploying seat) with $20/month usage credit shared across all Vercel usage. 10,000 concurrent sandboxes, 24-hour sessions, max 8 vCPU / 16 GB, vCPU allocation up to 5,000 per minute) - Free tier: Hobby: free monthly allowance of 5 Active CPU hours, 420 GB-hours of memory, 5,000 creations and 20 GB data transfer; no overage billing, creation pauses until the next cycle. - Isolation: Firecracker microVM per sandbox - Pause and resume: yes - Snapshots: yes - GPU: no - Custom images: yes - Network controls: yes - Max session hours: 24 - Max vCPU: 8 - Max memory GiB: 16 - SDKs: JavaScript/TypeScript, Python - Regions: 19 Vercel regions, default iad1 (Washington, D.C.); region chosen per sandbox or per project. Rates for CPU, memory, data transfer and drives vary by region. - Open source: no - Self-hostable: no - Best for: Agent loops that mostly wait on LLM or API calls, since idle CPU is not billed; Teams already on Vercel that want sandboxes under the same account, credit and regions; Workloads needing egress policies with domain allowlists, credential injection or deny-all, changeable on a live sandbox - Watch out for: Memory is billed for the full provisioned amount for the whole session, even when idle; Hobby sessions are capped at 45 minutes and usage stops (no overage) once the monthly allowance is used; Max 8 vCPU / 16 GB per sandbox on Pro; larger sizes need Enterprise; Persistent sandboxes create snapshots on every stop that are billed as storage until they expire or are deleted - Benchmark (ComputeSDK benchmarks, 2026-11-13): Burst start time (median of 100 at once) 428 ms, rank 9 of 27 - Benchmark (ComputeSDK benchmarks, 2026-11-13): Repo clone, install and typecheck (one run) 72265 ms, rank 20 of 25 - Benchmark (StarSling HPC sandbox benchmarks, 2026-09-29): Node.js web tooling (runs per second) 12.96 runs/s, rank 10 of 12 - Benchmark (StarSling HPC sandbox benchmarks, 2026-09-29): Better Auth cold dependency install 20.7 s, rank 10 of 11 - Sources: https://vercel.com/docs/vercel-sandbox/pricing https://vercel.com/docs/plans/pro-plan https://vercel.com/docs/vercel-sandbox https://vercel.com/docs/sandbox/concepts/persistent-sandboxes https://vercel.com/docs/sandbox/concepts/firewall - Facts checked: 2026-11-14 ## Mosaic Sandbox Mosaic Sandbox runs each sandbox as a Firecracker microVM on KVM, driven through a REST API, the mos CLI, Python, TypeScript and Go SDKs, MCP or native SSH. One price covers CPU, RAM and disk: $0.05 per active hour for each 2 vCPU / 4 GB billing unit, with larger shapes billed as multiple units. A quiet sandbox hibernates after three seconds and costs nothing until the next request wakes it, and egress has no per-GB charge. A task mode clones a repository, installs dependencies, runs a command and reports each step. - Page: https://aisandboxhunt.com/mosaic-sandbox - Company: Mosaic - Website: https://sandbox.mosaicos.com - Categories: AI agent sandboxes - Score: 74/100, #10 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $8.80 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Fixed-size price per hour: $0.05 (2 vCPU / 4 GB (one billing unit, the default shape)) - Billing increment: per second ($0.00001389 per second per unit) - Billed while: Only while the sandbox is awake. It hibernates after 3 quiet seconds and is not billed while hibernated; the next API, file, exec or SSH request wakes it. - Plans: Pay as you go $0.00/month (One SKU for everyone. Up to 1,000 sandboxes alive at once per organization (hibernated ones included). RBAC, SSO and audit logs included) - Free tier: No free tier or trial credit is stated on the pricing or signup pages. - Isolation: Firecracker microVM on Linux KVM per sandbox; no shared-root containers - Pause and resume: yes - Snapshots: yes - GPU: no - Custom images: yes - Network controls: yes - Max vCPU: 16 - Max memory GiB: 32 - SDKs: Python, JavaScript/TypeScript, Go, CLI, MCP server, REST API - Regions: Silicon Valley and Northern Virginia, placed automatically with failover or pinned per sandbox. No EU region today. - Open source: no - Self-hostable: no - Best for: Agents with long gaps between commands, since hibernation is free and wakes on the next request; Repository tasks that should clone, install and test in one call; Developers who want plain SSH access to a sandbox - Watch out for: A 1 vCPU sandbox costs the same as 2 vCPU / 4 GB; Only US regions today; No free tier or credit is stated; No GPU support - Sources: https://sandbox.mosaicos.com/pricing/ https://sandbox.mosaicos.com https://sandbox.mosaicos.com/llms.txt https://sandbox.mosaicos.com/start/ - Facts checked: 2026-11-13 ## Upstash Box Upstash Box gives each agent an isolated container with a shell, filesystem, git, snapshots, an optional managed Chromium and built-in coding agents such as Claude Code, Codex and OpenCode. Standard boxes bill only for CPU actually consumed, at $0.10 per core-hour, and freeze automatically when idle while keeping their storage; disk is $0.10 per GB-month. Keep-alive boxes stay on continuously for a flat $0.0111 to $0.0444 per hour by size. A free tier gives 10 concurrent boxes and 5 CPU hours a month. - Page: https://aisandboxhunt.com/upstash-box - Company: Upstash - Website: https://upstash.com/docs/box - Categories: AI agent sandboxes - Score: 73/100, #11 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $10.56 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.10 (active CPU only) - Price per GiB-hour of memory: $0.00 - Storage per GiB-month: $0.10 - Billing increment: by CPU time consumed, in core-hours - Billed while: Standard boxes: only for CPU actually consumed; idle boxes with no CPU use cost nothing for compute and freeze automatically. Storage bills while data exists. Keep-alive boxes bill a flat hourly rate for as long as they exist. - Plans: Free $0.00/month (Up to 10 concurrent boxes and 5 CPU hours per month. Agents use your own model provider key); Pay as You Go $0.00/month (Up to 1,000 concurrent boxes (soft limit, can be raised). Standard boxes at $0.10 per active CPU hour plus $0.10 per GB-month of storage; keep-alive boxes at a fixed hourly rate) - Free tier: Free plan: 5 CPU hours per month and up to 10 concurrent boxes. - Isolation: Isolated container per box with its own filesystem, shell and network stack - Pause and resume: yes - Snapshots: yes - Custom images: no - Network controls: yes - Max vCPU: 8 - Max memory GiB: 16 - SDKs: JavaScript/TypeScript, Python, MCP server - Regions: AWS us-east-1 only; more regions planned. - Open source: no - Self-hostable: no - Best for: Agents that mostly wait, since standard boxes bill only for CPU actually used; Always-on agent boxes at a low flat monthly price with keep-alive; Running Claude Code, Codex or OpenCode in a box with your own model key - Watch out for: Container isolation rather than a VM; Only one region (AWS us-east-1); No custom Docker images yet; Idle boxes freeze and stop their running processes - Sources: https://upstash.com/pricing/box https://upstash.com/docs/box/overall/how-it-works https://upstash.com/docs/box/guides/eve-setup - Facts checked: 2026-11-13 ## CreateOS Sandbox CreateOS Sandbox runs each workload in a Firecracker microVM with its own kernel, created from a TypeScript SDK, CLI, MCP server or REST API in EU or US regions. Every plan pays the same per-second rate: $0.03616363 per vCPU-hour, $0.01159025 per GiB-hour of memory and $0.0001584 per GB-hour of disk, with no egress charges. Pausing snapshots memory and device state and stops vCPU billing, but memory and storage keep billing while paused. Plans only set limits such as concurrent sandboxes and the largest size. - Page: https://aisandboxhunt.com/createos - Company: NodeOps, Inc. - Website: https://createos.sh/products/sandbox - Categories: AI agent sandboxes - Score: 72/100, #12 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $20.89 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.03616 - Price per GiB-hour of memory: $0.01159 - Storage per GiB-month: $0.1156 - Billing increment: per second - Billed while: vCPU, memory and disk while running. Paused sandboxes are still billed for memory and disk, but not vCPU. Disk bills for as long as the sandbox exists. - Plans: Free $0.00/month (1 concurrent sandbox, 10 per day, largest shape 1 vCPU / 1 GB, 10 GiB disk, no extra disks or templates. 500 credits at signup); Beginner $10.00/month (From $10 to $50 a month, with 1,000 to 5,000 initial credits and a 1.2x top-up bonus. 5 concurrent sandboxes, 50 per day, largest shape 4 vCPU / 4 GB, 30 GiB disk); Pro $75.00/month (From $75 to $200 a month, with 7,500 to 20,000 initial credits and a 1.5x top-up bonus. 20 concurrent sandboxes, 200 per day, largest shape 8 vCPU / 8 GB, 50 GiB disk) - Free tier: New accounts start with 500 free credits, no card required; the Free plan allows one sandbox at a time. - Isolation: Firecracker microVM per sandbox with its own guest kernel; egress rules enforced on the host outside the VM - Pause and resume: yes - Snapshots: yes - GPU: no - Custom images: yes - Network controls: yes - Max vCPU: 8 - Max memory GiB: 8 - SDKs: JavaScript/TypeScript, CLI, MCP server, REST API - Regions: EU and US regions, chosen per sandbox. Running on your own infrastructure is available to enterprise customers only. - Open source: no - Self-hostable: yes - Best for: Agents that sit idle between steps, since pause keeps full memory state and stops vCPU billing; Workloads that move a lot of data out, because egress is not charged; Multi-agent setups that need sandboxes talking over a private network - Watch out for: Paused sandboxes still pay for memory and disk; The Free plan runs one sandbox at a time, up to 1 vCPU / 1 GB; No Python SDK yet; Python goes through the REST API; Concurrent limits count paused sandboxes until they are destroyed - Sources: https://createos.sh/pricing/sandbox https://createos.sh/products/sandbox https://createos.sh/docs/Sandbox/Limits/ https://createos.sh/docs/Account-Billing/Pricing https://createos.sh/llms.txt - Facts checked: 2026-11-13 ## LangSmith Sandboxes LangSmith Sandboxes are microVMs for agent-generated code, created from the LangSmith Python or TypeScript SDK or a CLI. Agents get root, Docker and a default Ubuntu image with common languages and tools, or you boot from your own container-built snapshot. Compute, memory and storage are billed per second, and every plan includes 8 LSU (US$8) of sandbox usage each month. Sandboxes stop after an idle timeout, wake on the next command, and can be snapshotted and forked. - Page: https://aisandboxhunt.com/langsmith-sandboxes - Company: LangChain - Website: https://www.langchain.com/langsmith/sandboxes - Categories: AI agent sandboxes - Score: 72/100, #13 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $33.26 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.0576 - Price per GiB-hour of memory: $0.01845 - Storage per GiB-month: $0.08979 - Billing increment: per second - Billed while: Compute and memory while the sandbox runs; storage for persistent sandboxes while their filesystem is kept. Stopped sandboxes pay storage only. - Plans: Developer $0.00/month (1 seat. Includes 8 LSU (US$8) of sandbox usage per month; limited to 10 sandboxes); Plus $39.00/month ($39 per seat per month. Includes 8 LSU (US$8) of sandbox usage per month; usage beyond that is pay as you go) - Free tier: Every plan, including the free Developer plan, includes 8 LSU (US$8) of sandbox usage each month. Developer is limited to 10 sandboxes. - Isolation: MicroVM per sandbox - Pause and resume: yes - Snapshots: yes - GPU: no - Custom images: yes - Network controls: yes - Max vCPU: 16 - Max memory GiB: 64 - SDKs: Python, TypeScript, CLI - Regions: LangSmith Cloud regions (US or EU); also available on BYOC and self-hosted deployments. - Open source: no - Self-hostable: yes - Best for: Teams already using LangSmith or Deep Agents that want sandboxes in the same account; Agents that need root, Docker and a full toolchain in a disposable VM; Branching agent runs by snapshotting a sandbox and forking it - Watch out for: Outbound network access is restricted to an allowlist unless you request an exemption; The free Developer plan is limited to 10 sandboxes; Idle sandboxes stop after 10 minutes by default and stopped ones are deleted after 14 days; Billing is in LSU credits shared with other LangSmith products - Sources: https://www.langchain.com/pricing-langsmith https://www.langchain.com/langsmith/sandboxes https://docs.langchain.com/langsmith/sandbox-sdk https://docs.langchain.com/langsmith/sandboxes - Facts checked: 2026-11-13 ## Podflare Podflare runs each sandbox as a KVM-backed microVM with its own kernel, root access, outbound internet and a Python REPL whose state carries across calls. Sandboxes can be forked into copy-on-write children and frozen to disk as Persistent Spaces. Compute is metered per second at $0.05 per vCPU-hour and $0.016 per GiB-hour plus $0.09 per GB of egress, and monthly tiers (Free, Pro $50, Scale $200) set sandbox size, lifetime and concurrency. The service is labelled private alpha, and new accounts get $200 of credit. - Page: https://aisandboxhunt.com/podflare - Company: Podflare - Website: https://podflare.ai - Categories: AI agent sandboxes - Score: 71/100, #15 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $28.86 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.05 - Price per GiB-hour of memory: $0.016 - Billing increment: per second, billed monthly - Billed while: Pay-as-you-go compute for usage beyond the tier's base capacity; the pricing page does not say whether idle sandboxes inside the warm window are billed. - Plans: Free $0.00/month (1 GB RAM, 2 vCPU, 4 GB disk per sandbox; 5-minute idle timeout; 30-minute maximum lifetime; 10 concurrent sandboxes); Pro $50.00/month (4 GB RAM, 4 vCPU, 16 GB disk per sandbox; 30-minute idle timeout; 8-hour maximum lifetime; 50 concurrent sandboxes; 24-hour support response. Includes base capacity of unstated size); Scale $200.00/month (16 GB RAM, 8 vCPU, 64 GB disk per sandbox; 2-hour idle timeout; 24-hour maximum lifetime; 500 concurrent sandboxes; 4-hour support response. Includes base capacity of unstated size) - Free tier: New accounts get $200 of credit, which applies to both subscription and usage fees; no card hold during the private alpha. - Isolation: KVM-backed microVM per sandbox with a dedicated guest kernel - Pause and resume: yes - Snapshots: yes - Network controls: yes - Max session hours: 24 - Max vCPU: 8 - Max memory GiB: 16 - SDKs: Python, JavaScript/TypeScript, MCP server, REST API - Regions: The pricing page lists EU and US-West; the docs list five regions (us-west, us-central, us-east, eu, sg), routed to the nearest by default with optional region pinning. - Open source: no - Self-hostable: no - Best for: Tree-search or multi-branch agents that fork a running sandbox into many copies; Code-interpreter tools that need Python state kept across calls; Prototyping on a $200 starting credit - Watch out for: Private alpha: APIs and tier policy may still change, and paid upgrades are activated by email; Free tier sandboxes stop after 30 minutes; Scale tops out at 24 hours; Egress is billed at $0.09 per GB; SOC 2 is not yet certified - Sources: https://podflare.ai/pricing https://podflare.ai https://docs.podflare.ai/ - Facts checked: 2026-11-13 ## Agent37 Cloud Agent37 Cloud runs persistent instances for code, services and agent harnesses, managed through a REST API with a Python client. Default instances are gVisor-isolated containers on shared hosts in the United States; VM instances (own kernel, KVM) cost 2.5x the compute rate and also run in France and India. Rates are published per 730-hour month ($0.80 per vCPU, $0.70 per GB of RAM, $0.09 per GB of disk) and metered per minute: compute while running, disk while the instance is kept. There is no subscription; usage draws down a prepaid balance. - Page: https://aisandboxhunt.com/agent37 - Company: Agent37 - Website: https://agent37.com - Categories: AI agent sandboxes, Dev boxes for AI agents - Score: 70/100, #16 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $1.06 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.001096 - Price per GiB-hour of memory: $0.0009589 - Storage per GiB-month: $0.09 - Billing increment: per minute - Billed while: Compute (vCPU and RAM) for the whole time an instance is running, including starting and other transitional states. Sleeping or stopped instances bill disk only. - Plans: Prepaid wallet $0.00/month (No subscription or minimum; top-ups of $5 to $10,000. The limit is the memory running at once: 32 GB before the first top-up, 40 GB after it, 200 GB once top-ups total $50, 800 GB at $250 and 3,200 GB at $1,000. Sleeping and stopped instances do not count. Max 4 vCPU and 32 GB RAM per instance) - Free tier: One-time $5 starter credit after adding a card (the card is not charged), for new accounts; up to 3 instances running at once until the first top-up. - Isolation: gVisor per default instance; optional VM instances with their own kernel and KVM hardware virtualization - Pause and resume: yes - Snapshots: yes - Custom images: yes - Network controls: yes - Max vCPU: 4 - Max memory GiB: 32 - SDKs: REST API, Python - Regions: Default instances run in the United States. VM instances run in us, eu (France) and ap-south (India, Mumbai); outside the US only VM instances are available. - Open source: no - Self-hostable: no - Best for: One persistent instance per end user or agent that sits idle most of the day, using auto-sleep to pay mostly for disk; Hosting agent harnesses such as Claude Code, Codex or OpenCode as always-on instances; Small, steady workloads on a prepaid balance with no plan fee - Watch out for: Default instances are gVisor containers on shared hosts; full VMs cost 2.5x compute; Maximum 4 vCPU per instance on every type; Default instances run only in the United States; other regions require VM instances; Running instances at once are capped by total top-ups, starting at 3 - Benchmark (ComputeSDK benchmarks, 2026-11-13): Burst start time (median of 100 at once) 3832 ms, rank 23 of 27 - Benchmark (ComputeSDK benchmarks, 2026-11-13): Repo clone, install and typecheck (one run) 193271 ms, rank 25 of 25 - Sources: https://www.agent37.com/pricing https://www.agent37.com/docs/agents-api/billing https://www.agent37.com/docs/agents-api/instance-sizing https://www.agent37.com/docs/agents-api/vm-instances https://www.agent37.com/docs/agents-api/concepts https://agent37.com https://www.agent37.com/docs/agents-api/instance-auto-sleep https://www.agent37.com/docs/llms.txt - Facts checked: 2026-11-15 ## CodeSandbox CodeSandbox runs VM Sandboxes as microVMs that use memory snapshotting to start, resume and fork in seconds, either from the CodeSandbox SDK or as cloud development environments in its browser editor and VS Code. Runtime is billed in VM credits by machine size, from 5 credits ($0.075) an hour for Pico to 320 credits ($4.80) an hour for XLarge, metered per minute. The free Build plan includes 400 credits a month and 10 concurrent VMs; the Scale plan from $170 a month raises concurrency to 250 VMs and sizes to 16 vCPU. CodeSandbox is part of Together AI. - Page: https://aisandboxhunt.com/codesandbox - Company: CodeSandbox (part of Together AI) - Website: https://codesandbox.io - Categories: AI agent sandboxes, Dev boxes for AI agents - Score: 69/100, #17 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $26.40 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Fixed-size price per hour: $0.075 (Pico: 1 core / 2 GB (5 credits per hour)) - Billing increment: per minute (rounded up) - Billed while: VM runtime while a VM Sandbox is running; hibernated sandboxes are resumed from memory snapshots. - Plans: Build $0.00/month (400 VM credits a month (about 40 hours of Nano), 5 members, 10 concurrent VM Sandboxes, 20 new SDK sandboxes an hour, 1,000 SDK requests an hour, VMs up to 4 vCPU / 8 GiB, 20 GB storage per VM); Scale $170.00/month (From $170 a month per workspace. Up to 20 members, 250 concurrent VMs, 1,000 new sandboxes an hour, 10,000 requests an hour, VMs up to 16 vCPU / 32 GiB, on-demand credits. The included VM credits are stated as 160 hours a month on the pricing page and 1,100 credits in the SDK docs) - Free tier: Build plan: free, with 400 VM credits each month (unused credits do not roll over) and up to 10 concurrent VMs. - Isolation: MicroVM per VM Sandbox, with memory snapshots for resume and fork - Pause and resume: yes - Snapshots: yes - Max vCPU: 16 - Max memory GiB: 32 - SDKs: JavaScript/TypeScript - Regions: Not stated on the pages checked. - Open source: no - Self-hostable: no - Best for: Products that embed a browser IDE or per-user dev environments; Workloads that resume or fork running VMs from memory snapshots; Trying an SDK sandbox for free with a monthly credit allowance - Watch out for: The free plan caps VMs at 4 vCPU / 8 GiB and 10 concurrent; Billing rounds up to whole minutes; Prices and included credits differ between the pricing page and the docs; Only a JavaScript/TypeScript SDK is documented - Sources: https://codesandbox.io/pricing https://codesandbox.io/docs/sdk/pricing https://codesandbox.io/docs/learn/credit-usage/credits https://codesandbox.io/docs/learn/vm-sandboxes/overview - Facts checked: 2026-11-14 ## Superserve Superserve runs each sandbox in a Firecracker microVM started from a template, driven by TypeScript and Python SDKs, a CLI, a REST API or an MCP server. Compute ($0.0504 per vCPU-hour) and allocated memory ($0.0162 per GiB-hour) bill per second while a sandbox is active; pausing checkpoints memory, processes and files and stops compute billing, and storage bills for retained data. Runtime is unlimited and a paused sandbox is kept until deleted unless an auto-delete window is set. Paid plans are monthly fees that include the same amount of usage. - Page: https://aisandboxhunt.com/superserve - Company: Superserve - Website: https://superserve.ai - Categories: AI agent sandboxes, Dev boxes for AI agents - Score: 69/100, #19 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $29.15 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.0504 - Price per GiB-hour of memory: $0.0162 - Storage per GiB-month: $0.07884 - Billing increment: per second - Billed while: Compute and allocated memory while the sandbox is active; nothing for compute while paused. Storage (shared template once plus each sandbox's retained changes) bills for as long as it is kept, including paused sandboxes. - Plans: Developer $20.00/month (Includes $20 of usage. Up to 100 concurrent sandboxes, unlimited runtime, up to 4 vCPU and 16 GiB memory per sandbox. Auto Reload adds $5 when the balance drops under $2. Email support); Pro $250.00/month (Includes $250 of usage; extra usage billed at the end of each billing period. Up to 2,000 concurrent sandboxes, unlimited runtime, up to 16 vCPU and 64 GiB memory per sandbox. Priority Slack and email support) - Free tier: Start free with no card: a new account gets a limited free trial for its first sandbox (the credit amount is not stated). Trial sandboxes can use up to 2 vCPU and 4 GiB memory. After the trial credit is used, a card and a plan are required. - Isolation: Firecracker microVM per sandbox - Pause and resume: yes - Snapshots: yes - Custom images: yes - Network controls: yes - Max vCPU: 16 - Max memory GiB: 64 - SDKs: TypeScript, Python, CLI, REST API, MCP server - Regions: Not stated on the pages checked. - Open source: no - Self-hostable: no - Best for: Long-running agent sessions that pause and resume with memory, processes and files intact; Branching work: snapshot a sandbox and fork copies to try several approaches in parallel; Agents that need API credentials without the key entering the sandbox, plus egress allow and deny rules - Watch out for: vCPU and allocated memory bill for the whole time a sandbox is active, busy or idle; pause it to stop compute charges; Sandbox size is fixed by its template; a different size needs a different template; The free trial amount is not published, and paid use needs a monthly plan; No region choice or GPU support is documented on the pages checked - Benchmark (ComputeSDK benchmarks, 2026-11-13): Burst start time (median of 100 at once) 377 ms, rank 8 of 27 - Benchmark (ComputeSDK benchmarks, 2026-11-13): Repo clone, install and typecheck (one run) 57422 ms, rank 9 of 25 - Benchmark (aisandboxbench, 2026-10-23): Cold start (median) 0.52 s, rank 1 of 9 - Benchmark (aisandboxbench, 2026-10-23): Agent tool-call replay (median) 13.147 s, rank 2 of 9 - Sources: https://www.superserve.ai/pricing/ https://docs.superserve.ai/introduction https://docs.superserve.ai/sandbox/lifecycle https://docs.superserve.ai/templates/build-spec https://docs.superserve.ai/sandbox/create - Facts checked: 2026-11-14 ## Alibaba Cloud Agent Sandbox Agent Sandbox is Alibaba Cloud's isolated execution environment for AI agents, with code, browser, desktop and mobile sandbox types. The fully managed Flash Edition works with the E2B SDK and CLI, while the Container Edition exposes sandboxes through the Kubernetes API on Container Compute Service. Each sandbox runs in its own VM environment, and a newer MicroVM runtime in beta adds snapshots, forking and Windows. Billing is per second for vCPU, memory and disk above a 15 GiB free allowance, in three compute classes; idle sandboxes can drop into light or deep hibernation, which stops CPU or CPU and memory charges. - Page: https://aisandboxhunt.com/alibaba-agent-sandbox - Company: Alibaba Cloud - Website: https://www.alibabacloud.com/help/en/agent-sandbox/product-overview/agent-sandbox-overview - Categories: AI agent sandboxes - Score: 68/100, #20 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $12.57 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.01786 - Price per GiB-hour of memory: $0.008928 - Storage per GiB-month: $0.1847 - Billing increment: per second, 1-second minimum, billed hourly - Billed while: While active: vCPU, memory and disk above 15 GiB, on the size set at creation. Light hibernation (Flash Edition only): memory and disk, no vCPU. Deep hibernation and saved snapshots: storage only, for memory size plus disk size. Nothing after the sandbox is destroyed. - Plans: Pay as you go $0.00/month (Three compute classes (Economy, Default, Performance) with the same billing rules and different unit prices. Flash Edition default quota of 300 sandboxes per region) - Free tier: No free trial or monthly allowance is stated for Agent Sandbox; only the first 15 GiB of disk on an active sandbox is free. - Isolation: Each sandbox runs in an independent VM environment; the second-generation runtime (beta) uses lightweight MicroVM isolation - Pause and resume: yes - Snapshots: yes - Custom images: yes - Network controls: yes - Max session hours: 24 - SDKs: E2B SDK (Python, TypeScript), E2B CLI, Java and Go SDKs maintained by Alibaba Cloud, Kubernetes API (Container Edition), Alibaba Cloud OpenAPI - Regions: Flash Edition: China (Beijing, Ulanqab, Shanghai, Hangzhou, Shenzhen, Hong Kong), Singapore, Malaysia (Johor), US (Virginia) and US (Silicon Valley). Container Edition: over 25 regions across China, Asia-Pacific, Europe, the Americas and the Middle East. - Open source: no - Self-hostable: no - Best for: Teams already on Alibaba Cloud, or serving users in China and Asia-Pacific, that need agent sandboxes; Moving E2B SDK code to another provider by changing the endpoint and API key; Long agent sessions that wait on models or people, since hibernation stops CPU and memory charges; Reinforcement learning rollouts and evaluations that clone many sandboxes from one checkpoint - Watch out for: Pause and resume need an allowlisted account, and snapshots are an invited test; The MicroVM runtime that adds snapshots and Windows is still beta and only accepts Ubuntu- or Debian-based images; Per-sandbox CPU and memory limits are not published; Prices differ between mainland China and other regions, and the Performance class is only in the Flash Edition - Sources: https://www.alibabacloud.com/help/en/agent-sandbox/product-overview/agent-sandbox-overview https://www.alibabacloud.com/help/en/agent-sandbox/product-overview/billing-overview https://www.alibabacloud.com/help/en/agent-sandbox/product-overview/limits-of-agent-sandbox https://www.alibabacloud.com/help/en/agent-sandbox/product-overview/open-service-area https://www.alibabacloud.com/help/en/agent-sandbox/pause-and-resume https://www.alibabacloud.com/help/en/agent-sandbox/snapshots https://www.alibabacloud.com/help/en/agent-sandbox/runtime-gen1-vs-gen2 https://www.alibabacloud.com/help/en/agent-sandbox/network-access-control https://www.alibabacloud.com/help/en/agent-sandbox/build-sandbox-templates-with-ghcr-custom-images https://www.alibabacloud.com/help/en/agent-sandbox/e2b-sdk-java-and-go https://www.alibabacloud.com/help/en/functioncompute/pay-as-you-go-of-fc-agent-sandbox - Facts checked: 2026-11-15 ## Tencent CubeSandbox CubeSandbox is an open-source sandbox service built on RustVMM and KVM that runs each sandbox in its own microVM with a dedicated kernel. It accepts the E2B SDK, so clients can switch by changing one environment variable, and supports pause and resume, snapshots, cloning and rollback, eBPF egress filtering and a credential-injecting proxy. It deploys on a single node, bare metal, cloud VMs, Kubernetes (preview) or Tencent Cloud via Terraform, and includes a web console. - Page: https://aisandboxhunt.com/cubesandbox - Company: Tencent Cloud - Website: https://github.com/TencentCloud/CubeSandbox - Categories: AI agent sandboxes - Score: 68/100, #21 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: Self-host · no list price for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Billing increment: Not applicable: free software run on your own infrastructure - Billed while: Not applicable: you pay only for your own hardware or cloud machines - Free tier: Free open-source software, licensed Apache-2.0 apart from listed third-party components. - Isolation: MicroVM per sandbox with a dedicated guest kernel (RustVMM on KVM), plus eBPF network isolation between sandboxes - Pause and resume: yes - Snapshots: yes - GPU: no - Custom images: yes - Network controls: yes - SDKs: E2B SDK (Python, JavaScript/TypeScript), REST API, Web console - Regions: Wherever you deploy it: x86_64 Linux with KVM on bare metal or cloud VMs, Kubernetes, or Tencent Cloud via Terraform. - Open source: yes (Apache-2.0) - Self-hostable: yes - Best for: Moving E2B SDK code to sandboxes on your own servers without changing client code; Dense deployments with many small microVM sandboxes per host; Agents that need checkpoints, rollback or forking of a running sandbox - Watch out for: The quick-start path needs an x86_64 Linux host with KVM; No GPU support yet; it is on the roadmap; Kubernetes deployment and cross-node resume are still in preview - Sources: https://github.com/TencentCloud/CubeSandbox https://github.com/TencentCloud/CubeSandbox/blob/master/LICENSE https://github.com/TencentCloud/CubeSandbox/releases/tag/v0.8.0 - Facts checked: 2026-11-13 ## E2B E2B runs each sandbox as a Firecracker microVM started from a template, controlled through JavaScript/TypeScript and Python SDKs. Running sandboxes are billed per second for allocated vCPU and RAM; disk is included and paused sandboxes do not incur compute charges. A paused sandbox keeps its filesystem and memory indefinitely and resets the continuous-runtime limit (1 hour on Hobby, 24 hours on Pro). The runtime is open source under Apache-2.0, and Enterprise customers can deploy into their own AWS, Google Cloud or Azure account (BYOC). - Page: https://aisandboxhunt.com/e2b - Company: E2B - Website: https://e2b.dev - Categories: AI agent sandboxes - Score: 68/100, #22 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $29.15 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.0504 - Price per GiB-hour of memory: $0.0162 - Billing increment: per second - Billed while: Wall-clock while the sandbox is running, for allocated vCPU and RAM; compute billing stops when the sandbox is paused, killed or times out. - Plans: Hobby $0.00/month ($100 one-time usage credit (not monthly). Sessions up to 1 hour of continuous runtime, 10 concurrent sandboxes, up to 8 vCPU and 8 GiB RAM, 10 GiB disk, 1 sandbox created per second); Pro $150.00/month (No usage credit included. Sessions up to 24 hours of continuous runtime, 100 concurrent sandboxes (600 with a +$500/month add-on, 1,100 with a +$1,000/month add-on), 8+ vCPU and 8+ GiB RAM, 20+ GiB disk, 5 sandboxes created per second) - Free tier: Hobby plan: $0/month with a one-time $100 usage credit, no credit card required; 1-hour sessions and 10 concurrent sandboxes. - Isolation: Firecracker microVM per sandbox - Pause and resume: yes - Snapshots: yes - GPU: no - Custom images: yes - Network controls: yes - Max session hours: 24 - Max vCPU: 8 - Max memory GiB: 8 - SDKs: JavaScript/TypeScript, Python - Regions: Shared E2B cloud; a shared EU cluster is available on every plan on request for workspaces with a payment method; BYOC on AWS, Google Cloud and Azure for Enterprise. - Open source: yes - Self-hostable: yes - Best for: Agent sessions that pause between turns: pausing keeps filesystem and memory and stops compute billing; Teams that want an open-source (Apache-2.0) runtime they can inspect or run themselves; Workloads that need egress allow/deny lists by domain or CIDR, changeable on a running sandbox - Watch out for: No GPU support; Pro costs $150/month and includes no usage credit; Hobby sandboxes stop after 1 hour of continuous runtime unless paused and resumed; Per-sandbox size tops out at 8 vCPU / 8 GiB on Hobby; larger sizes require Pro or contacting E2B - Benchmark (ComputeSDK benchmarks, 2026-11-13): Burst start time (median of 100 at once) 1280 ms, rank 18 of 27 - Benchmark (ComputeSDK benchmarks, 2026-11-13): Repo clone, install and typecheck (one run) 64219 ms, rank 15 of 25 - Benchmark (aisandboxbench, 2026-10-23): Cold start (median) 0.793 s, rank 4 of 9 - Benchmark (aisandboxbench, 2026-10-23): Agent tool-call replay (median) 15.468 s, rank 3 of 9 - Benchmark (StarSling HPC sandbox benchmarks, 2026-09-29): Node.js web tooling (runs per second) 12.61 runs/s, rank 11 of 12 - Benchmark (StarSling HPC sandbox benchmarks, 2026-09-29): Better Auth cold dependency install 20.2 s, rank 8 of 11 - Sources: https://e2b.dev/pricing https://e2b.dev/docs/billing https://e2b.dev/docs/sandbox/persistence https://e2b.dev/docs/sandbox/internet-access https://e2b.dev/docs/faq https://github.com/e2b-dev/infra - Facts checked: 2026-11-14 ## E2B open-source runtime E2B Runtime is the full backend that powers E2B Cloud: the control-plane API, the per-node orchestrator that drives Firecracker, the in-VM agent, the edge router and the template builder. Sandboxes are restored from snapshots, can be paused, resumed and forked, and get per-sandbox egress allow and deny lists. The single-machine Embed package runs the whole stack on one Linux host with KVM via Docker Compose, and also ships as Terraform for GCP, AWS and Azure and as a Kubernetes manifest. The software is free; multi-node production deployments in your own cloud are offered by E2B as an enterprise service without a published price. - Page: https://aisandboxhunt.com/e2b-runtime - Company: E2B - Website: https://github.com/e2b-dev/runtime - Categories: AI agent sandboxes - Score: 68/100, #23 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: Self-host · no list price for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Billing increment: Not applicable: free software run on your own infrastructure - Billed while: Not applicable: you pay only for your own hardware or cloud resources - Free tier: Free open-source software under the Apache-2.0 license. E2B's hosted cloud and dedicated enterprise deployments are separate offerings. - Isolation: One Firecracker microVM per sandbox, in its own cgroup and network namespace - Pause and resume: yes - Snapshots: yes - Custom images: yes - Network controls: yes - SDKs: JavaScript/TypeScript, Python, CLI, REST API - Regions: Wherever you deploy it: one Linux host with KVM, or GCP, AWS or Azure via Terraform, or Kubernetes. - Open source: yes (Apache-2.0) - Self-hostable: yes - Best for: Teams that want the same sandbox stack locally or on their own servers that E2B runs in its cloud; Agent workloads that need pause, resume and forking of running microVMs; Existing E2B SDK users who need sandboxes inside their own account - Watch out for: The free Embed package runs on a single machine; multi-node setups are an enterprise offering; Needs Linux hosts with KVM; Several backing services to operate: PostgreSQL, Redis, ClickHouse and object storage - Sources: https://github.com/e2b-dev/runtime https://github.com/e2b-dev/runtime/releases/tag/2026.30 - Facts checked: 2026-11-13 ## microsandbox microsandbox runs untrusted code in small hardware-isolated virtual machines on your own machine, started from the msb CLI or embedded directly in an application through its SDKs, with no server or daemon to run. Sandboxes boot from standard OCI container images, can be forked while running, snapshotted and restored, and can be limited to allowed hosts with secrets that never enter the VM. It is free Apache-2.0 software and the project describes itself as beta. - Page: https://aisandboxhunt.com/microsandbox - Company: Super Rad Company - Website: https://github.com/superradcompany/microsandbox - Categories: AI agent sandboxes - Score: 68/100, #24 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: Self-host · no list price for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Billing increment: Not applicable: free software run on your own hardware - Billed while: Not applicable: you pay only for your own hardware or cloud machines - Free tier: Free open-source software under the Apache-2.0 license; no hosted paid tier with a published price was found. - Isolation: Hardware-isolated microVM per sandbox (KVM on Linux, Apple Silicon virtualization on macOS, Windows Hypervisor Platform on Windows) - Pause and resume: yes - Snapshots: yes - Custom images: yes - Network controls: yes - SDKs: TypeScript, Python, Rust, Go, Ruby, CLI, MCP server - Regions: Runs on your own hardware: Linux with KVM, macOS on Apple Silicon, or Windows with Windows Hypervisor Platform. - Open source: yes (Apache-2.0) - Self-hostable: yes - Best for: Running agent or user code in microVMs on a laptop or single server without a separate daemon; Embedding sandbox creation directly in TypeScript, Python, Rust, Go or Ruby applications; Workflows that fork or snapshot a running sandbox and restore it later - Watch out for: Beta software: breaking changes and missing features are expected; Needs hardware virtualization: KVM on Linux, Apple Silicon on macOS, or Windows Hypervisor Platform; You run and scale the hosts yourself; there is no multi-node scheduler described - Sources: https://github.com/superradcompany/microsandbox https://github.com/superradcompany/microsandbox/releases/tag/v0.7.15 - Facts checked: 2026-11-13 ## Solo Enterprise for Agent Substrate Solo.io's enterprise release of Agent Substrate and kagent, generally available since November 2026, runs AI agents on a Kubernetes cluster you operate. Agent Substrate is an open-source runtime, co-maintained by Solo.io and Google, that runs every agent session as an actor in its own gVisor or microVM sandbox on a pool of pre-started worker pods. When a turn ends it snapshots the actor's files and memory to object storage and frees the worker, then restores the actor on any free worker when the next request arrives. kagent adds declarative agents, harnesses for Claude Code, Codex and custom images, short-lived standalone sandboxes for scripts, and an egress gateway that injects credentials so the sandbox holds none. The enterprise build adds single sign-on, access control, tracing, air-gapped installs and support; it needs license keys and is priced by quote. - Page: https://aisandboxhunt.com/solo-agent-substrate - Company: Solo.io - Website: https://www.solo.io/products/kagent - Categories: AI agent sandboxes - Score: 68/100, #25 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: Self-host · no list price for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Billing increment: Not applicable: enterprise license by quote; no usage rates are published - Billed while: Not applicable: software runs on your own cluster, and you pay your provider for its nodes and storage - Free tier: The open-source kagent and Agent Substrate projects are free to run yourself. The enterprise release, with support, single sign-on, access control, tracing and air-gapped installs, needs license keys priced by quote from Solo.io sales. - Isolation: gVisor sandbox per actor by default, or a microVM under Cloud Hypervisor with a Kata Containers kernel; many sandboxed actors share one worker pod, and plain containers are not supported - Pause and resume: yes - Snapshots: yes - Custom images: yes - Network controls: yes - Max session hours: 24 - SDKs: kagent CLI, Kubernetes API (kubectl/YAML), MCP server, A2A - Regions: Any Kubernetes cluster you run: 1.34 or later in certificate compatibility mode, 1.37 or later in the default mode, including OpenShift. Air-gapped installs are supported. - Open source: yes (Apache-2.0) - Self-hostable: yes - Best for: Platform teams that want many mostly idle agent sessions on the Kubernetes clusters they already run; Regulated or air-gapped environments that need a supported, self-hosted agent runtime; Running untrusted agent code without giving the sandbox API keys, with default-deny outbound traffic - Watch out for: Self-hosted only: you operate the cluster, PostgreSQL, object storage and upgrades; Enterprise pricing is by quote only, and the install needs separate license keys; Recent Kubernetes is required (1.34 or later), and there is no in-place upgrade from the older 0.5.x kagent line - Sources: https://www.solo.io/press-releases/announcing-kagent-1-0-ga https://www.solo.io/blog/announcing-kagent-1-0-ga https://www.solo.io/pricing https://docs.solo.io/kagent/latest/about/architecture/agent-substrate/ https://docs.solo.io/kagent/latest/reference/changelog/release-notes/ https://docs.solo.io/kagent/latest/about/system-requirements/ https://docs.solo.io/kagent/latest/about/substrate-runtime/sandboxing/ https://docs.solo.io/kagent/latest/about/substrate-runtime/standalone-sandboxes/ https://docs.solo.io/kagent/latest/setup/licensing/ https://www.solo.io/discover/agent-substrate https://github.com/agent-substrate/substrate/releases/tag/v1.0.0 - Facts checked: 2026-11-15 ## Tenki Sandbox Tenki Sandbox runs each session in a hardware-isolated Linux VM with its own kernel, root access and Docker support, driven from a CLI, an MCP server or Go, TypeScript and Python SDKs. Sessions bill per second at $0.0504 per CPU core-hour and $0.0162 per GiB-hour of memory, with the first 5 GiB of disk free. Sessions can be paused with memory intact, snapshotted (disk and memory) and forked. Tenki also sells GitHub Actions runners from the same account and credits. - Page: https://aisandboxhunt.com/tenki - Company: Tenki Cloud (Luxor Technology) - Website: https://tenki.cloud/products/sandbox - Categories: AI agent sandboxes, CI runners - Score: 67/100, #27 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $29.15 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.0504 - Price per GiB-hour of memory: $0.0162 - Storage per GiB-month: $0.20 - Billing increment: per second - Billed while: CPU cores and memory while a session is active; a paused session stops compute charges. Root disk above 5 GiB bills per second; snapshots and volumes bill from the workspace storage pool. - Plans: Starter $0.00/month (5 members, 5 concurrent sessions, up to 4 cores and 8 GB RAM per VM, 100 GiB storage. One-time $100 free credit once a payment method is added; usage beyond it billed at the listed rates); Team $250.00/month ($250 a month billed monthly or $200 a month billed yearly; $100 of credit each month. 50 members, 100 concurrent sessions, up to 16 cores and 64 GB RAM per VM, 1,000 GiB storage, priority support) - Free tier: One-time $100 in free credits the first time a payment method is added; nothing is credited before a card is on file. Adding the card also turns on auto-reload, which adds $20 whenever the balance drops below $5, and can be turned off. - Isolation: Hardware-isolated VM per session with its own kernel - Pause and resume: yes - Snapshots: yes - Custom images: yes - Network controls: yes - Max vCPU: 16 - Max memory GiB: 64 - SDKs: Go, JavaScript/TypeScript, Python, CLI, MCP server - Regions: Not stated. Runs on Tenki-owned hardware with a Ceph storage cluster. - Open source: no - Self-hostable: no - Best for: Agents that need a full VM with root, systemd and Docker; Long agent sessions that pause with memory intact and resume later; Teams that also want cheaper GitHub Actions runners on the same bill - Watch out for: Starter caps each VM at 4 cores / 8 GB and 5 active sessions; Snapshot and volume storage costs $0.20 per GiB-month from the first byte; Regions are not documented; Templates do not accept Dockerfiles - Sources: https://tenki.cloud/docs/pricing https://tenki.cloud/docs/account/free-credits https://tenki.cloud/products/sandbox https://tenki.cloud/docs/sandbox/snapshots https://tenki.cloud/docs/sandbox/sessions - Facts checked: 2026-11-14 ## AWS Lambda MicroVMs Lambda MicroVMs launch Firecracker microVMs from a snapshot of your own Dockerfile-built image, each with its own HTTPS endpoint. You pick a baseline size (vCPU scales with memory at 1 vCPU per 2 GB) and the VM can burst to 4x that baseline, paying the baseline while running plus actual use above it, per second. Idle MicroVMs can be suspended with memory and disk preserved, paying only snapshot storage, and a MicroVM lives at most 8 hours. - Page: https://aisandboxhunt.com/aws-lambda-microvms - Company: Amazon Web Services - Website: https://aws.amazon.com/lambda/lambda-microvms/ - Categories: AI agent sandboxes - Score: 65/100, #29 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $44.39 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.0997 - Price per GiB-hour of memory: $0.0132 - Storage per GiB-month: $0.08 - Billing increment: per second - Billed while: While running: the configured baseline vCPU and memory for the whole time, plus actual use above the baseline when bursting. Suspended MicroVMs pay snapshot storage only; terminated ones pay nothing. - Free tier: No MicroVMs free allowance is listed. New AWS customers receive up to $200 in AWS Free Tier credits usable across AWS. - Isolation: Firecracker microVM per MicroVM - Pause and resume: yes - Snapshots: yes - GPU: no - Custom images: yes - Network controls: yes - Max session hours: 8 - Max vCPU: 16 - Max memory GiB: 32 - SDKs: AWS SDKs, AWS CLI, Lambda API - Regions: Multiple AWS Regions (account memory quotas are listed per Region); x86 is available in all MicroVMs Regions except a few in Asia Pacific and Europe. - Open source: no - Self-hostable: no - Best for: Interactive sandboxes that sit idle between user turns, since suspended VMs pay only storage; Per-tenant isolation with a dedicated HTTPS endpoint for each VM; Teams already on AWS that want VPC access from sandboxes - Watch out for: Hard 8-hour maximum lifetime per MicroVM, including suspended time; Every launch, suspend and resume adds per-GB snapshot read or write charges; Images must be built from a Dockerfile zip uploaded to S3 and snapshotted first; Baseline vCPU and memory are billed for the full running time, even when idle - Sources: https://aws.amazon.com/lambda/pricing/ https://aws.amazon.com/lambda/lambda-microvms/ https://docs.aws.amazon.com/lambda/latest/dg/microvms-images.html https://docs.aws.amazon.com/lambda/latest/dg/gettingstarted-limits.html https://docs.aws.amazon.com/lambda/latest/dg/lambda-microvms-guide.html https://aws.amazon.com/bedrock/agentcore/pricing/ - Facts checked: 2026-11-13 ## Daytona Daytona sandboxes are isolated environments with their own kernel, filesystem and network stack, offered as container, Linux VM, Windows, macOS and GPU sandbox classes. Billing is per second for the vCPU, RAM and disk reserved while a sandbox is started; stopped or paused sandboxes are billed for disk only, and archived sandboxes are not billed. There is no monthly plan fee: organizations move through usage tiers (Tier 1 to 4) by verifying and topping up credit, which raises resource pools, rate limits and network freedom. SDKs cover Python, TypeScript, Ruby, Go and Java. - Page: https://aisandboxhunt.com/daytona - Company: Daytona - Website: https://www.daytona.io - Categories: AI agent sandboxes, Dev boxes for AI agents, GPU sandboxes - Score: 64/100, #30 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $29.15 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.0504 - Price per GiB-hour of memory: $0.0162 - GPU prices per hour: Nvidia B300 (on-demand) $6.25, Nvidia B200 (on-demand) $6.25, AMD MI355X (on-demand) $5.99, Nvidia H200 (on-demand) $4.54, Nvidia H100 (on-demand) $3.95, Nvidia RTX PRO 6000 (on-demand) $3.03, Nvidia L40S (on-demand) $1.99, Nvidia RTX 5090 (on-demand) $1.29, Nvidia RTX 4090 (on-demand) $0.99, Nvidia B300 (preemptible) $4.08, Nvidia B200 (preemptible) $3.59, AMD MI355X (preemptible) $3.44, Nvidia H200 (preemptible) $2.61, Nvidia H100 (preemptible) $2.27, Nvidia RTX PRO 6000 (preemptible) $1.74, Nvidia L40S (preemptible) $1.14, Nvidia RTX 5090 (preemptible) $0.74, Nvidia RTX 4090 (preemptible) $0.57 - Storage per GiB-month: $0.07884 - Billing increment: per second - Billed while: Wall-clock for reserved resources: vCPU, RAM and disk while started (and while creating, starting, stopping or pausing); disk only while stopped or paused; nothing while archived or deleted. - Plans: Pay as you go (Tier 1-4) $0.00/month (No plan fee. Org-wide pools by tier: Tier 1 (email verified) 10 vCPU / 10-20 GiB RAM (docs conflict) / 30 GiB disk; Tier 2 (card + $25 top-up) 100 / 200 / 300; Tier 3 ($500 top-up) 250 / 500 / 2,000; Tier 4 ($2,000 top-up every 30 days) 500 / 1,000 / 5,000. Per sandbox: max 4 vCPU, 8 GiB RAM, 10 GiB disk (non-GPU). Tier 1-2 cannot override the default network restrictions) - Free tier: $200 in free compute on sign-up, no credit card required; free credits cannot be used for GPU sandboxes. Startups can apply for up to $50k in credits. - Isolation: Linux containers by default, each with its own kernel, filesystem and network stack; Linux VM and Windows sandbox classes run as full virtual machines and add pause, resume and fork. - Pause and resume: yes - Snapshots: yes - GPU: yes - Custom images: yes - Network controls: yes - Max vCPU: 4 - Max memory GiB: 8 - SDKs: Python, TypeScript, Ruby, Go, Java - Regions: Shared regions: United States (us) and Europe (eu). GPU sandboxes run in a global 'earth' region. Dedicated regions and custom (bring-your-own-compute) regions are also offered. - Open source: no - Self-hostable: no - Best for: Long-lived agent workspaces that can be stopped or archived between uses to cut cost to disk-only or zero; Agents that need GPU sandboxes (H100, H200, B200, B300, RTX, MI355X) with on-demand or cheaper preemptible pricing; Teams wanting SDKs beyond Python/TypeScript (Ruby, Go, Java) or Windows sandboxes - Watch out for: Default per-sandbox size limit is 4 vCPU / 8 GiB RAM / 10 GiB disk for non-GPU sandboxes; Tier 1 and Tier 2 organizations cannot change network restrictions per sandbox; full internet access needs Tier 3 ($500 top-up); Billing is for reserved resources, not usage: an idle started sandbox pays the full vCPU and RAM rate; The public GitHub repository stopped being maintained in June 2026; core development moved to a private codebase - Benchmark (ComputeSDK benchmarks, 2026-11-13): Burst start time (median of 100 at once) 682 ms, rank 12 of 27 - Benchmark (aisandboxbench, 2026-10-23): Cold start (median) 1.25 s, rank 8 of 9 - Benchmark (aisandboxbench, 2026-10-23): Agent tool-call replay (median) 12.274 s, rank 1 of 9 - Benchmark (StarSling HPC sandbox benchmarks, 2026-09-29): Node.js web tooling (runs per second) 21.52 runs/s, rank 3 of 12 - Benchmark (StarSling HPC sandbox benchmarks, 2026-09-29): Better Auth cold dependency install 11.8 s, rank 2 of 11 - Sources: https://www.daytona.io/pricing https://www.daytona.io/docs/en/billing https://www.daytona.io/docs/en/limits https://www.daytona.io/docs/en/sandboxes https://www.daytona.io/docs/en/network-limits https://www.daytona.io/docs/en/regions https://github.com/daytonaio/daytona - Facts checked: 2026-11-14 ## Blaxel Blaxel Sandboxes are microVMs for running agent and LLM-generated code, with a REST API, an MCP server and SDKs for files and processes. A sandbox without an active connection moves to standby after a few seconds; its memory, filesystem and running processes are snapshotted and resume in about 25 ms. Compute is billed per GB of allocated memory per second only while the sandbox is active; CPU is allocated from the memory size and not billed separately. Standby sandboxes pay only snapshot storage. - Page: https://aisandboxhunt.com/blaxel - Company: Blaxel - Website: https://blaxel.ai - Categories: AI agent sandboxes - Score: 63/100, #31 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $29.15 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per GiB-hour of memory: $0.0414 - Storage per GiB-month: $0.20 - Billing increment: per second (rate published per GB of RAM per second) - Billed while: allocated memory while the sandbox is active (has an active connection); standby time is not billed for compute, only snapshot storage - Plans: Pay-as-you-go (Tier 0) $0.00/month (Up to 10 sandboxes; sandbox TTL capped at 7 days; free credits of up to $200 (one-time, not monthly); community support); Pay-as-you-go (Tier 1) $20.00/month (Requires $20 of top-ups in the trailing 30 days; the top-up is prepaid credit for usage, not a fee. Up to 50 sandboxes; sandbox TTL capped at 30 days); Pay-as-you-go (Tier 2) $50.00/month (Requires $50 of top-ups in the trailing 30 days (prepaid credit for usage). Up to 200 sandboxes; higher tiers (up to Tier 9, 100,000+ sandboxes) are listed in the console) - Free tier: Up to $200 in free credits at sign-up; Tier 0 quotas (10 sandboxes, 7-day maximum sandbox lifetime). Promotional credits do not count toward tier upgrades. - Isolation: microVM (hardware-isolated, one per sandbox; hypervisor not named) - Pause and resume: yes - Snapshots: yes - GPU: no - Custom images: yes - Network controls: yes - SDKs: TypeScript, Python, Go, CLI, REST API, MCP server - Regions: us-pdx-1 (Oregon), us-was-1 (North Virginia), eu-lon-1 (London), eu-fra-1 (Frankfurt); not all regions are self-service by default - Open source: no - Self-hostable: no - Best for: Agents that sit idle between turns and must resume quickly with processes intact; Many long-lived per-user sandboxes kept in standby at storage-only cost; Workloads that need domain allowlists or secret injection on outbound traffic - Watch out for: Any open connection or keep-alive keeps the sandbox active and billing for its full memory allocation; Tier 0 and Tier 1 sandboxes are deleted after at most 7 and 30 days; longer persistence needs a higher top-up tier; Writable filesystem lives in RAM (about 50% of memory), so disk-heavy work needs more memory or a volume; Manual snapshots and forking are in private preview on the Carbon generation and not recommended for production - Benchmark (ComputeSDK benchmarks, 2026-11-13): Burst start time (median of 100 at once) 351 ms, rank 7 of 27 - Benchmark (ComputeSDK benchmarks, 2026-11-13): Repo clone, install and typecheck (one run) 60706 ms, rank 12 of 25 - Benchmark (StarSling HPC sandbox benchmarks, 2026-09-29): Node.js web tooling (runs per second) 19.48 runs/s, rank 6 of 12 - Benchmark (StarSling HPC sandbox benchmarks, 2026-09-29): Better Auth cold dependency install 12.1 s, rank 3 of 11 - Sources: https://blaxel.ai/pricing https://docs.blaxel.ai/Sandboxes/Overview https://docs.blaxel.ai/Sandboxes/Expiration https://docs.blaxel.ai/Infrastructure/Regions https://docs.blaxel.ai/Sandboxes/Fork - Facts checked: 2026-11-14 ## Render Sandboxes Render Sandboxes run each sandbox in its own Firecracker microVM, created and driven from the Render CLI, the TypeScript or Python SDK, or the API. CPU is billed at $0.12 per active CPU hour and memory at a fixed hourly rate for the compute plan's RAM while the sandbox runs; a suspended sandbox has no compute charge. By default a sandbox suspends after 15 minutes without commands or file transfers and resumes on the next one, and both the idle and runtime timeouts can be turned off so a sandbox runs indefinitely. The product is in public beta and runs in one region (Oregon). - Page: https://aisandboxhunt.com/render - Company: Render - Website: https://render.com/sandboxes - Categories: AI agent sandboxes, Dev boxes for AI agents - Score: 63/100, #32 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $26.05 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.12 (active CPU only) - Price per GiB-hour of memory: $0.019 - Storage per GiB-month: $0.08 - Billing increment: prorated to the second (Render's general compute billing rule; no separate increment stated for sandboxes) - Billed while: CPU: only active CPU time while the sandbox runs. Memory: the compute plan's full provisioned RAM while the sandbox runs. Nothing for compute while suspended; snapshot storage billed separately. - Plans: Hobby $0.00/month (No workspace fee; sandbox compute billed at usage rates. 100 concurrent sandboxes by default, 1 seat, no maximum session length); Pro $25.00/month ($25/month workspace fee plus compute. 1,000 concurrent sandboxes by default (up to 5,000 on request), unlimited seats, no maximum session length); Scale $499.00/month ($499/month workspace fee plus compute. 1,000 concurrent sandboxes by default (up to 10,000 on request), multiple workspaces, SAML SSO, no maximum session length) - Free tier: No free sandbox usage is stated. The Hobby workspace has no monthly fee, but sandbox compute is billed at the usage rates; Render's free compute plans cover web services, Key Value and Postgres only. - Isolation: Firecracker microVM per sandbox - Pause and resume: yes - Snapshots: yes - Custom images: yes - Network controls: yes - Max vCPU: 8 - Max memory GiB: 16 - SDKs: TypeScript, Python, CLI, REST API - Regions: 1 region during the beta: Oregon. - Open source: no - Self-hostable: no - Best for: Agent work that waits on model or API calls, since idle CPU is not billed and idle sandboxes suspend; Teams already hosting on Render that want sandboxes in the same workspace and billing; Long-lived agent environments: timeouts can be turned off and suspended sandboxes resume with their state - Watch out for: Public beta, with one region (Oregon) and limits that may change; Memory is billed for the plan's full RAM while the sandbox runs, even when idle; Sizes are fixed plans up to 8 CPU / 16 GB; no GPU is described; No public URLs for apps running inside a sandbox during the beta - Sources: https://render.com/pricing https://render.com/sandboxes https://render.com/docs/sandboxes https://render.com/docs/sandboxes-limits https://render.com/docs/sandboxes-creating - Facts checked: 2026-11-14 ## Docker Sandboxes Docker Sandboxes run AI coding agents in microVMs on your own machine, each with its own Docker daemon, created and managed with the sbx CLI. All outbound traffic passes through a host proxy that enforces network policy and injects credentials so the agent never sees real keys. The CLI and local compute are free, including for commercial use, but the software is not open source and requires a Docker sign-in. Cloud sandboxes on Docker-managed compute are billed per second through a separate pay-as-you-go subscription whose rates are not published in the docs. - Page: https://aisandboxhunt.com/docker-sandboxes - Company: Docker - Website: https://docs.docker.com/ai/sandboxes/ - Categories: AI agent sandboxes - Score: 62/100, #33 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: Self-host · no list price for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Billing increment: Local: not applicable. Cloud: metered per second (rates not published) - Billed while: Local: free. Cloud: by sandbox runtime and the vCPU and memory configuration - Free tier: The sbx CLI and local sandbox compute are free, including for commercial work. Cloud sign-ups get a one-time promotional credit (amount not stated). - Isolation: MicroVM per sandbox (hypervisor isolation) with its own Docker daemon - Pause and resume: yes - GPU: yes - Network controls: yes - SDKs: CLI (sbx), Sandboxes API and SDK (cloud) - Regions: Local sandboxes run on your machine: macOS 14+ on Apple Silicon, Windows 11 x64 with Windows Hypervisor Platform, or Ubuntu 24.04+ with KVM. Cloud sandbox regions not stated. - Open source: no - Self-hostable: yes - Best for: Running Claude Code, Codex and other coding agents on a laptop with VM isolation and full Docker access; Keeping API credentials outside the agent's environment through proxy injection; Teams already standardized on Docker tooling - Watch out for: Not open source and requires a Docker sign-in; Cloud sandbox prices are not published in the docs; Each sandbox runs a full VM and its own Docker daemon, so overhead is higher than a container - Sources: https://docs.docker.com/ai/sandboxes/ https://docs.docker.com/ai/sandboxes/architecture/ https://docs.docker.com/ai/sandboxes/install/ https://docs.docker.com/agentic-platform/signup/ https://docs.docker.com/subscription-billing/plans/docker-agentic-platform/ https://github.com/docker/sbx-releases/releases/tag/v0.53.0 - Facts checked: 2026-11-13 ## Runloop Runloop runs agent workloads in Devboxes, VM-isolated Linux sandboxes created through Python and TypeScript SDKs or a REST API. Compute is billed per second for CPU and memory while a Devbox is initializing, running, suspending or resuming, with separate per-GB-hour charges for Devbox, blueprint, snapshot and object storage. The Basic plan has no monthly fee; Pro costs $250/month and adds suspend/resume; new accounts get a Pro trial with $50 of usage credit. - Page: https://aisandboxhunt.com/runloop - Company: Runloop - Website: https://runloop.ai - Categories: AI agent sandboxes, Dev boxes for AI agents - Score: 62/100, #36 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $55.76 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.108 - Price per GiB-hour of memory: $0.0252 - Storage per GiB-month: $0.2499 - Billing increment: per second - Billed while: Compute (CPU and memory) while the Devbox status is initializing, running, suspending or resuming; storage keeps billing while suspended. - Plans: Basic $0.00/month (Free subscription plus usage-based compute; 100 GB free storage; Devboxes, blueprints and snapshots; email support. Suspend/resume is listed as a Pro feature); Pro $250.00/month ($250/month plus usage; 1 TB free storage; adds suspend and resume, repo connections, beta features and Slack support) - Free tier: New accounts start on a Pro trial with a one-time $50 usage credit and no credit card (3 concurrent Devboxes, sizes up to MEDIUM, 1-hour max keep-alive). After that, the Basic plan has no monthly fee and bills usage only. - Isolation: Virtual machine (microVM) per Devbox - Pause and resume: yes - Snapshots: yes - Custom images: yes - Network controls: yes - Max vCPU: 16 - Max memory GiB: 64 - SDKs: Python, TypeScript, REST API - Regions: Not published on the pages checked; Enterprise can deploy into the customer's own VPC. - Open source: no - Self-hostable: yes - Best for: Coding-agent pipelines that need snapshots, reusable blueprint images and suspend/resume; Long-lived but intermittent agent sessions on Pro, where suspended Devboxes stop compute billing; Teams that need network egress policies and an Enterprise VPC deployment option - Watch out for: CPU costs $0.108/CPU-hour and is charged on the allocated size while running, not on active use; Suspend/resume is a Pro feature, and Pro costs $250/month on top of usage; Devboxes default to a 1-hour lifetime; the trial caps keep-alive at 1 hour and sizes at MEDIUM; Suspend keeps only disk state; processes must be restarted after resume - Benchmark (StarSling HPC sandbox benchmarks, 2026-09-29): Node.js web tooling (runs per second) 8.29 runs/s, rank 12 of 12 - Benchmark (StarSling HPC sandbox benchmarks, 2026-09-29): Better Auth cold dependency install 24.8 s, rank 11 of 11 - Sources: https://www.runloop.ai/pricing https://docs.runloop.ai/docs/devboxes/configuration/sizes https://docs.runloop.ai/docs/devboxes/start-stop https://docs.runloop.ai/docs/overview/your-runloop-trial https://docs.runloop.ai/llms.txt - Facts checked: 2026-11-14 ## DigitalOcean Agent Harness Runtime Harness Runtime is part of DigitalOcean Managed Agents, in public preview for all users. Each session runs in a hardware-isolated microVM that can host a coding agent such as Claude Code or Codex, or act as a plain sandbox for running commands. CPU is billed on actual consumption and memory on allocation, per second, and paused sessions stop compute charges while keeping files and history. Checkpoints capture the whole machine so you can fork or roll back. - Page: https://aisandboxhunt.com/digitalocean-harness-runtime - Company: DigitalOcean - Website: https://docs.digitalocean.com/products/managed-agents/ - Categories: AI agent sandboxes - Score: 61/100, #38 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $50.00 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.044 (active CPU only) - Price per GiB-hour of memory: $0.0095 - Storage per GiB-month: $0.05 - Billing increment: per second - Billed while: While a session is active: CPU on actual consumption (zero while the agent waits), memory on the allocated size over time. Paused sessions pay no compute; volumes, checkpoints and snapshots keep billing as storage. - Plans: Agent Droplets Pro $50.00/month ($50/month fee funds a balance spent at a 15% discount, covering $58.82 of eligible usage at standard rates. Usage beyond that bills at standard rates. Unlimited agents subject to fair use); Agent Droplets Team $200.00/month ($200/month fee funds a balance spent at a 20% discount, covering $250 of eligible usage at standard rates. Usage beyond that bills at standard rates. Unlimited agents subject to fair use) - Free tier: New customers receive a one-time $5 promotional credit for Managed Agents usage. - Isolation: Hardware-isolated microVM per session - Pause and resume: yes - Snapshots: yes - GPU: no - Custom images: yes - Network controls: yes - Max vCPU: 16 - Max memory GiB: 32 - SDKs: doctl CLI, REST API, PyDo (Python), Control panel, Cuttlefish iOS app - Regions: RIC1 (Richmond) only, per the availability table. - Open source: no - Self-hostable: no - Best for: Running hosted coding agents such as Claude Code, Codex or OpenCode in isolated machines; Agent work that waits a lot on models or people, since idle CPU is not billed and sessions can pause; Branching work from a checkpoint to try several approaches in parallel - Watch out for: Public preview, and only one region (RIC1) is listed; Memory is billed on the allocated size even while the agent is idle; No private VPC access, live app preview or lifecycle webhooks yet; Requires prepaid funds or a plan before sessions can start - Sources: https://docs.digitalocean.com/products/managed-agents/agent-harness-runtime/details/pricing/ https://docs.digitalocean.com/products/managed-agents/agent-harness-runtime/how-to/pay-for-harness-runtime/ https://docs.digitalocean.com/products/managed-agents/agent-harness-runtime/details/features/ https://docs.digitalocean.com/products/managed-agents/agent-harness-runtime/details/limits/ https://docs.digitalocean.com/products/managed-agents/ https://docs.digitalocean.com/products/managed-agents/agent-harness-runtime/details/availability/ - Facts checked: 2026-11-13 ## AWS Bedrock AgentCore Code Interpreter AgentCore Code Interpreter is a managed tool in Amazon Bedrock AgentCore that runs Python, JavaScript and TypeScript in isolated sessions for AI agents. CPU is billed only for actual consumption, so I/O wait costs nothing, while memory is billed on the peak used so far, both per second. Sessions default to 15 minutes and can be extended to 8 hours, with Sandbox or Public network modes and S3 access through an execution role. - Page: https://aisandboxhunt.com/aws-agentcore-code-interpreter - Company: Amazon Web Services - Website: https://aws.amazon.com/bedrock/agentcore/ - Categories: AI agent sandboxes - Score: 60/100, #40 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $16.10 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.0895 (active CPU only) - Price per GiB-hour of memory: $0.00945 - Billing increment: per second, 1-second minimum; 128 MB minimum memory billing - Billed while: From microVM boot to session termination. CPU only for actual consumption (no charge during I/O wait); memory for the peak consumed up to each second, including idle periods. - Free tier: No Code Interpreter free allowance. New AWS customers receive up to $200 in AWS Free Tier credits usable across AWS. - Isolation: Isolated session per sandbox; AWS describes billing from microVM boot to microVM shutdown - GPU: no - Network controls: yes - Max session hours: 8 - SDKs: AWS SDKs (including Python/boto3), AWS CLI, AgentCore API - Regions: Not stated on the pages checked; available in AWS Regions where AgentCore is offered. - Open source: no - Self-hostable: no - Best for: Agents already built on Amazon Bedrock or AgentCore that need a code tool; Workloads with long I/O waits, since idle CPU is not billed; Data analysis on files stored in Amazon S3 - Watch out for: Only Python, JavaScript and TypeScript runtimes are provided; Memory is billed on peak use for the whole session, even while idle; Per-session CPU and memory limits are not published; Data transfer is billed separately at EC2 rates - Sources: https://aws.amazon.com/bedrock/agentcore/pricing/ https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/code-interpreter-tool.html https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/code-interpreter-resource-session-management.html - Facts checked: 2026-11-13 ## Railway Sandboxes Railway Sandboxes are ephemeral Linux VMs scoped to a Railway environment, created from the TypeScript SDK, CLI, dashboard or GraphQL API. They boot Ubuntu with git, Node and common coding agents preinstalled, and can join the project's private network to reach its databases and services. CPU is billed for seconds actually used and memory for what is in use, so an idle sandbox still pays for memory. Checkpoints save the disk to boot or fork later. - Page: https://aisandboxhunt.com/railway-sandboxes - Company: Railway - Website: https://railway.com/sandboxes - Categories: AI agent sandboxes - Score: 59/100, #42 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $56.22 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.06944 (active CPU only) - Price per GiB-hour of memory: $0.06944 - Billing increment: per second - Billed while: While the sandbox runs: CPU for active use only, memory in use (including OS and file cache) even while idle, and outbound traffic. Destroying the sandbox stops compute charges. - Plans: Free $0.00/month ($1 of usage credit per month. 10 running sandboxes per environment, max 2 vCPU / 2 GB per sandbox, idle timeout up to 5 minutes); Hobby $5.00/month (Includes $5 of usage per month. 50 running sandboxes per environment, max 8 vCPU / 8 GB, idle timeout up to 2 hours or disabled); Pro $20.00/month (Includes $20 of usage per month. 100 running sandboxes per environment, max 24 vCPU / 24 GB (the pricing page says 32 vCPU / 32 GB), idle timeout up to 2 hours or disabled) - Free tier: Free plan with $1 of usage credit each month, after a 30-day trial with a one-time $5 credit. Sandboxes are available on Free and the trial, capped at 2 vCPU / 2 GB. - Isolation: Isolated Linux VM per sandbox - Snapshots: yes - GPU: no - Custom images: yes - Network controls: yes - Max vCPU: 24 - Max memory GiB: 24 - SDKs: TypeScript, CLI, GraphQL API, Dashboard - Regions: US West (default), US East, EU West and Southeast Asia. - Open source: no - Self-hostable: no - Best for: Agents that need to reach a project's existing databases and services over a private network; Short tasks with low average CPU use, since only busy CPU seconds are billed; Trying several agent attempts from one prepared checkpoint - Watch out for: Memory in use keeps billing while a sandbox sits idle; Sandbox CPU and memory rates are well above Railway's normal service rates; Free and trial sandboxes are capped at 2 vCPU / 2 GB with a 5-minute idle timeout; No pause: checkpoints restore files into a new sandbox, not running processes - Sources: https://railway.com/sandboxes https://railway.com/pricing https://docs.railway.com/reference/pricing/plans - Facts checked: 2026-11-13 ## Kata Containers Kata Containers is a building block, not a ready-made sandbox service: a container runtime that starts each container or Kubernetes pod in its own lightweight virtual machine, so workloads get VM isolation while keeping container tooling. It supports several hypervisors and runs on x86_64, ARM64, IBM Power and IBM Z. Kubernetes sandbox projects use it as a runtime class. It is free Apache-2.0 software. - Page: https://aisandboxhunt.com/kata-containers - Company: Kata Containers community (OpenInfra) - Website: https://katacontainers.io - Categories: AI agent sandboxes - Score: 58/100, #43 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: Self-host · no list price for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Billing increment: Not applicable: free software run on your own hardware - Billed while: Not applicable: no usage charges - Free tier: Free open-source software under the Apache-2.0 license. - Isolation: Lightweight VM per container or pod, using a choice of hypervisors with hardware virtualization - Custom images: yes - SDKs: OCI runtime, containerd shim, Kubernetes RuntimeClass - Regions: Runs on your own 64-bit hosts with hardware virtualization (x86_64, ARM64, ppc64le, s390x). - Open source: yes (Apache-2.0) - Self-hostable: yes - Best for: Giving Kubernetes pods VM-level isolation through a runtime class; Running untrusted container workloads with existing container tooling; Platforms that need IBM Power or IBM Z support - Watch out for: Building block only: no sandbox API or SDK of its own; Requires hardware virtualization on the host (or nested virtualization in cloud VMs); More moving parts than a plain container runtime: hypervisor, guest kernel and agent - Sources: https://github.com/kata-containers/kata-containers https://github.com/kata-containers/kata-containers/releases/tag/4.3.1 - Facts checked: 2026-11-13 ## kosa8 Cloud kosa8 runs each agent in a microVM with its own kernel, either on your own Mac or Linux machines or on kosa8 Cloud, its hosted service in the US (Dallas, Texas) and the EU (Helsinki). Cloud sandboxes are created from Python, JavaScript/TypeScript, Go, Ruby or Java SDKs or a REST API and bill per second at $0.048 per vCPU-hour and $0.0154 per GB-hour of RAM. Idle sandboxes go to sleep and then cost only $0.08 per GB-month for their snapshot. Cloud is pay-as-you-go with a card on file and no free credit; the paid kosa8 plans ($29 to $199 a month) add features for your own machines and include no Cloud credit. - Page: https://aisandboxhunt.com/kosa8 - Company: Kosa8 Labs - Website: https://kosa8.com - Categories: AI agent sandboxes - Score: 58/100, #44 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $27.74 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.048 - Price per GiB-hour of memory: $0.0154 - Storage per GiB-month: $0.08 - Billing increment: per second - Billed while: While a sandbox runs. An idle sandbox sleeps on a timer you set and then pays only for its snapshot until the next command wakes it. - Plans: Cloud pay as you go $0.00/month (No plan needed; pay only for Cloud usage. A payment method is required before anything runs. 20 sandboxes running or paused per organization (sleeping ones do not count)); Solo $29.00/month (4 forks at once, 64 GB sandboxes; overage $0.10 per fork-hour if turned on); Pro $89.00/month (16 forks at once, 128 GB sandboxes); Team $169.00/month (Per person per month, from 3 people. 48 shared forks at once, 256 GB sandboxes, governance and MCP policies); Scale $199.00/month (Per person per month, from 5 people. 160 shared forks at once, 500 GB sandboxes) - Free tier: No Cloud credit of any kind: a card is required and usage is billed per second from the first second. Until its first payment, a new organization can run up to $50 of usage a month. Local sandboxes on your own computer are free. - Isolation: Hardware-virtualised microVM per sandbox with its own kernel (Cloud Hypervisor on KVM on Linux) - Pause and resume: yes - Snapshots: yes - GPU: no - SDKs: Python, JavaScript/TypeScript, Go, Ruby, Java, REST API, MCP server, CLI - Regions: US (us-ash1, Dallas, Texas) and EU (eu-hel1, Helsinki), at the same price. A sandbox goes to the region nearest the caller unless you name one or set a default region for the organization. Asia-Pacific is planned at 1.2x the price. You can also join your own servers as hosts. - Open source: no - Self-hostable: yes - Best for: Teams that want the same sandboxes on laptops, their own servers and a hosted cloud; Agents that idle between commands, since sleeping sandboxes pay only snapshot storage; EU data residency on the hosted service, by setting the organization's default region to Helsinki - Watch out for: Hosted sandboxes go to the nearest region, US or EU, unless you set a default region for the organization; A card is required before any Cloud sandbox runs; Maximum sandbox size on Cloud is not published; Several security features shown on the site (egress broker, undo) apply to local use and are not yet on Cloud; Plans add no Cloud credit; all Cloud usage is billed - Sources: https://docs.kosa8.com/guides/cloud/ https://kosa8.com/pricing/ https://kosa8.com/faq/ - Facts checked: 2026-11-15 ## Novita Agent Sandbox Novita Agent Sandbox gives each AI agent an isolated runtime with its own filesystem, processes and network, driven from Python and JavaScript SDKs or a CLI. Compute is billed per second for allocated vCPU ($0.03528 per vCPU-hour) and memory ($0.01152 per GiB-hour) while a sandbox runs; paused sandboxes keep files and memory and pay only for storage above a 60 GB account allowance. New accounts start on a free tier capped at 5 concurrent sandboxes, and a one-time $100 promotional credit is available after a short setup survey. - Page: https://aisandboxhunt.com/novita-sandbox - Company: Novita AI - Website: https://novita.ai/sandbox - Categories: AI agent sandboxes - Score: 58/100, #45 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $20.53 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.03528 - Price per GiB-hour of memory: $0.01152 - Storage per GiB-month: $0.0657 - Billing increment: per second - Billed while: Allocated vCPU and memory while the sandbox is running. Paused sandboxes pay no compute, only persistent storage above the 60 GB account allowance. - Plans: Free tier $0.00/month (Default for new accounts: 5 concurrent sandboxes, 1-hour maximum session, up to 2 vCPU and 4 GiB per sandbox, no priority scheduling. Usage is paid from account balance or promotional credit); Paid tier (pay as you go) $0.00/month (Unlocks automatically once the account balance is above $0: 100 concurrent sandboxes, up to 8 vCPU and 8 GiB per sandbox, priority scheduling. Session cap is 3 hours in the docs (24 hours on the product page); long-running mode lifts it) - Free tier: One-time $100 promotional credit for new users who complete an account setup survey, no credit card required. Credits may have a validity period. - Isolation: Separate runtime per sandbox with its own filesystem, process space and network stack; the underlying technology is not stated - Pause and resume: yes - Snapshots: yes - Custom images: yes - Network controls: yes - Max vCPU: 8 - Max memory GiB: 8 - SDKs: Python, JavaScript/TypeScript, CLI - Regions: Two US regions: us-phx-1 (default, supports secrets and snapshots) and us-virginia-1 (legacy). - Open source: no - Self-hostable: no - Best for: Agent tasks that need to pause with memory intact and resume later without paying for idle compute; Locked-down runs with outbound allowlists, private URLs and secrets injected outside the sandbox; Trying a sandbox API with a $100 credit and no card - Watch out for: Maximum size is 8 vCPU and 8 GiB per sandbox on the paid tier; Session limits differ between the product page (24 hours) and the docs (3 hours); The isolation technology is not documented; Running sandboxes are terminated when the account balance runs out - Sources: https://novita.ai/sandbox https://novita.ai/docs/guides/sandbox-pricing https://novita.ai/docs/guides/sandbox-faq-max-runtime https://novita.ai/docs/guides/sandbox-faq-cpu-memory https://novita.ai/docs/guides/sandbox-faq-isolation https://novita.ai/docs/guides/sandbox-faq-regions https://novita.ai/docs/guides/sandbox-snapshot - Facts checked: 2026-11-13 ## Microsoft NVX NVX is an ultra-light micro-VM sandbox for running untrusted workloads with hardware isolation, built on OpenVMM with a Linux guest. It was developed by Microsoft Research groups and runs on Linux with KVM or MSHV and on Windows with the Windows Hypervisor Platform. The nvx-rs command-line tool and Rust crate launch and supervise microVMs, with Alpine, Ubuntu and Azure Linux userlands included. Only development releases have been published so far. - Page: https://aisandboxhunt.com/microsoft-nvx - Company: Microsoft Research - Website: https://github.com/microsoft/nvx - Categories: AI agent sandboxes - Score: 57/100, #46 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: Self-host · no list price for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Billing increment: Not applicable: free software run on your own hardware - Billed while: Not applicable: no usage charges - Free tier: Free open-source software under the MIT license. - Isolation: Hardware-isolated micro-VM on OpenVMM with a Linux guest (KVM, MSHV or Windows Hypervisor Platform) - SDKs: Rust crate, CLI (nvx-rs) - Regions: Runs on your own Linux (KVM or MSHV) or Windows (WHP) host. - Open source: yes (MIT) - Self-hostable: yes - Best for: Building your own sandbox service on a small micro-VM that runs on both Linux and Windows hosts; Research and experiments with OpenVMM-based isolation; Rust projects that want to launch and supervise microVMs from code - Watch out for: Development releases only; no stable version yet; A building block: no hosted service, scheduler or SDKs beyond Rust; Needs KVM, MSHV or Windows Hypervisor Platform on the host - Sources: https://github.com/microsoft/nvx https://github.com/microsoft/nvx/releases https://github.com/microsoft/nvx/blob/dev/LICENSE - Facts checked: 2026-11-15 ## Modal Sandboxes Modal Sandboxes are containers created from the Python, JavaScript or Go SDK that run on either a gVisor runtime or a VM runtime with its own Linux kernel. They are billed per second at Modal's non-preemptible CPU and memory rates (3x the Function list price), charged on whichever is higher of the requested or actually used resources, so an unrequested sandbox pays mostly for what it uses. Sandboxes can attach GPUs (gVisor runtime only), run up to 24 hours, and save state through filesystem, directory and memory snapshots. Plans add a monthly compute credit: $30 on Starter and $100 on the $250/month Team plan. - Page: https://aisandboxhunt.com/modal - Company: Modal Labs - Website: https://modal.com - Categories: AI agent sandboxes, GPU sandboxes - Score: 57/100, #47 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $41.88 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.07096 - Price per GiB-hour of memory: $0.02401 - GPU prices per hour: Nvidia B300 $7.10, Nvidia B200 $6.25, Nvidia H200 SXM $4.54, Nvidia H100 SXM5 $3.95, Nvidia RTX PRO 6000 $3.03, Nvidia A100 80 GB $2.50, Nvidia A100 40 GB $2.10, Nvidia L40S $1.95, Nvidia A10 $1.10, Nvidia L4 $0.7992, Nvidia T4 $0.5904 - Storage per GiB-month: $0.09 - Billing increment: per second - Billed while: While the sandbox runs, charged on max(requested, actually used) CPU and memory; default request is 0.125 physical cores and 128 MiB. - Plans: Starter $0.00/month ($30/month compute credit, up to 3 seats, 100 containers and 10 GPU concurrency, 1 TiB/month free egress, 1-day log retention); Team $250.00/month ($100/month compute credit, unlimited seats, 5,000 containers and 50 GPU concurrency, 10 TiB/month free egress, custom domains for sandbox tunnels, static IP proxy) - Free tier: Starter plan: $0/month with $30 of compute credit every month (CPU, memory, GPU for Functions and Sandboxes). - Isolation: gVisor container runtime, or a VM runtime with its own Linux kernel (selectable per sandbox) - Pause and resume: no - Snapshots: yes - GPU: yes - Custom images: yes - Network controls: yes - Max session hours: 24 - SDKs: Python, JavaScript/TypeScript, Go - Regions: Runs in Modal's multi-cloud fleet; optional region pinning (us, eu, ap and narrower regions such as us-west, uk, jp) at a 1.15x (broad) or 1.75x (narrow) price multiplier. - Open source: no - Self-hostable: no - Best for: Bursty agent workloads where average CPU and memory use is far below peak, since billing follows actual usage when no request is set; Sandboxes that need GPUs, from T4 up to B300; Teams already running Python, JS or Go workloads on Modal who want sandboxes on the same account and credit - Watch out for: Sandbox CPU and memory cost 3x Modal's Function list rates; No pause/resume: sandboxes end after at most 24 hours and state must be carried over with snapshots; GPU sandboxes can be preempted and do not support memory snapshots; Region pinning adds a 1.15x to 1.75x price multiplier - Benchmark (ComputeSDK benchmarks, 2026-11-13): Burst start time (median of 100 at once) 451 ms, rank 10 of 27 - Benchmark (ComputeSDK benchmarks, 2026-11-13): Repo clone, install and typecheck (one run) 48147 ms, rank 4 of 25 - Benchmark (StarSling HPC sandbox benchmarks, 2026-09-29): Node.js web tooling (runs per second) 15.54 runs/s, rank 8 of 12 - Benchmark (StarSling HPC sandbox benchmarks, 2026-09-29): Better Auth cold dependency install 14.8 s, rank 5 of 11 - Sources: https://modal.com/pricing https://modal.com/docs/guide/resources https://modal.com/docs/guide/sandboxes https://modal.com/docs/guide/sandbox-resources https://modal.com/docs/guide/region-selection https://modal.com/docs/guide/sandbox-networking https://modal.com/docs/guide/sandbox-snapshots - Facts checked: 2026-11-14 ## Cloudflare Sandbox SDK Cloudflare's Linux sandboxes are Containers instances that a Worker starts through a Durable Object; each instance is a microVM with its own kernel running an image you provide. Billing follows Containers pricing: active CPU time plus provisioned memory and disk, metered every 10 ms while the instance runs, on top of the $5/month Workers Paid plan. Instances stop after an inactivity timeout of up to 6 hours, and files can be carried to the next instance with container snapshots (public beta). - Page: https://aisandboxhunt.com/cloudflare-sandbox - Company: Cloudflare - Website: https://developers.cloudflare.com/sandbox/ - Categories: AI agent sandboxes - Score: 56/100, #48 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $18.26 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.072 (active CPU only) - Price per GiB-hour of memory: $0.009 - Billing increment: per 10 ms - Billed while: While the container instance runs: CPU only for active usage; memory and disk for the provisioned instance size, even when idle. - Plans: Workers Paid $5.00/month (Includes 375 vCPU-minutes, 25 GiB-hours of memory and 200 GB-hours of disk per month (allowances, not a dollar credit). Default per-instance max 4 vCPU, 12 GiB memory, 20 GB disk; account max 15,000 concurrent vCPU and 60 TiB concurrent memory) - Free tier: None for sandboxes: Containers require the Workers Paid plan ($5/month), which includes 375 vCPU-minutes, 25 GiB-hours of memory and 200 GB-hours of disk each month. - Isolation: microVM per container instance (own kernel and network) - Pause and resume: no - Snapshots: yes - GPU: no - Custom images: yes - Network controls: yes - Max vCPU: 4 - Max memory GiB: 12 - SDKs: TypeScript (Workers / @cloudflare/sandbox) - Regions: Runs on Cloudflare's network; the pages checked do not document choosing a region. - Open source: no - Self-hostable: no - Best for: Apps already on Cloudflare Workers that need a full Linux environment per user or task; Bursty agent workloads that can stop instances quickly, since CPU is billed only when active; Running untrusted code with internet access turned off or outbound traffic routed through the Worker - Watch out for: Requires the $5/month Workers Paid plan; no free tier for Containers; Idle running instances still pay for provisioned memory and disk; Default instance ceiling is 4 vCPU and 12 GiB memory; No in-memory pause: instances stop after an inactivity timeout of up to 6 hours, and snapshots (public beta) restore files but not processes - Benchmark (ComputeSDK benchmarks, 2026-11-13): Repo clone, install and typecheck (one run) 107110 ms, rank 24 of 25 - Sources: https://developers.cloudflare.com/containers/platform/pricing/ https://developers.cloudflare.com/workers/platform/pricing/ https://developers.cloudflare.com/containers/platform/limits/ https://developers.cloudflare.com/sandbox/ https://developers.cloudflare.com/sandbox/concepts/lifetime/ - Facts checked: 2026-11-14 ## Hopx Hopx runs each sandbox in a Firecracker microVM created from a template, driven from Python or JavaScript SDKs, a CLI, a REST API or an MCP server. Sandboxes have no execution time limit unless you set a timeout, and can be paused with memory kept or stopped with the filesystem kept. Usage is billed per second: $0.000014 per vCPU-second and $0.0000045 per GiB-second of memory, plus storage. New accounts get $200 in free credits with no credit card. - Page: https://aisandboxhunt.com/hopx - Company: Bunnyshell - Website: https://hopx.ai - Categories: AI agent sandboxes, Dev boxes for AI agents - Score: 56/100, #49 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $29.15 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.0504 - Price per GiB-hour of memory: $0.0162 - Storage per GiB-month: $0.07884 - Billing increment: per second - Billed while: While the sandbox is running ('pay only when running'). The pricing section does not say whether paused or stopped sandboxes pay for storage. - Free tier: $200 in free credits for new accounts, no credit card required. The site does not say whether the credits expire. - Isolation: Firecracker microVM per sandbox - Pause and resume: yes - Custom images: yes - Network controls: yes - SDKs: Python, JavaScript/TypeScript, CLI, REST API, MCP server - Regions: Region chosen per sandbox; the docs show us-east and eu-west as examples and do not publish a full list. - Open source: no - Self-hostable: no - Best for: Running untrusted or model-generated code in microVMs from Python or JavaScript; Long-running agent or batch jobs, since sandboxes have no execution time limit by default; Desktop automation, with VNC, screenshots, mouse and keyboard control in the SDK - Watch out for: No published plans, concurrency limits or per-sandbox size limits; limits are said to depend on the account plan; Storage price is only given per GiB-second, and billing for paused or stopped sandboxes is not stated; Region list is not published; the docs only show us-east and eu-west as examples - Sources: https://hopx.ai https://docs.hopx.ai/core-concepts/sandboxes/managing-state https://docs.hopx.ai/core-concepts/sandboxes/timeout https://docs.hopx.ai/core-concepts/sandboxes/creating https://docs.hopx.ai/core-concepts/templates/configuration https://docs.hopx.ai/core-concepts/templates/building https://docs.hopx.ai/cli/commands/usage - Facts checked: 2026-11-14 ## Firecracker Firecracker is a building block, not a ready-made sandbox service: a minimal virtual machine monitor that uses Linux KVM to start lightweight microVMs, built at AWS for Lambda and Fargate. Each microVM is configured through a REST API on the host with its vCPUs, memory, network interfaces, disks and rate limiters. Kata Containers and many agent sandbox platforms build on it. It is free Apache-2.0 software. - Page: https://aisandboxhunt.com/firecracker - Company: Amazon Web Services - Website: https://firecracker-microvm.github.io - Categories: AI agent sandboxes - Score: 55/100, #51 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: Self-host · no list price for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Billing increment: Not applicable: free software run on your own hardware - Billed while: Not applicable: no usage charges - Free tier: Free open-source software under the Apache-2.0 license. - Isolation: KVM-based microVM per workload (Firecracker VMM) - Network controls: yes - SDKs: REST API (OpenAPI) - Regions: Runs on your own Linux hosts with KVM (bare metal or nested virtualization). - Open source: yes (Apache-2.0) - Self-hostable: yes - Best for: Building your own microVM sandbox or serverless platform; Workloads that need VM isolation with fast startup and small memory overhead; Rate-limiting network and disk bandwidth per VM - Watch out for: Building block only: no SDK, image tooling or orchestration included; Needs Linux hosts with KVM; Secure multi-tenant use depends on careful host configuration - Sources: https://github.com/firecracker-microvm/firecracker https://github.com/firecracker-microvm/firecracker/releases/tag/v1.18.0 - Facts checked: 2026-11-13 ## Promigence Promigence starts sandboxes from snapshots that must pass their own verify command at build time, and forks many identical copies at once for agent runs, evaluations, RL rollouts and CI jobs. Usage is metered per second at $0.0504 per vCPU-hour and $0.0162 per GiB-hour, paused time is not billed, and snapshot storage and egress are included. Every run is priced before it starts and held to a hard spend cap, and runs report environment failures separately from task failures. New accounts get $50 of free credit; paid plans are arranged by email. - Page: https://aisandboxhunt.com/promigence - Company: Promigence - Website: https://www.promigence.ai - Categories: AI agent sandboxes - Score: 55/100, #52 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $29.15 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.0504 - Price per GiB-hour of memory: $0.0162 - Storage per GiB-month: $0.00 - Billing increment: per second (1-second minimum) - Billed while: From hand-over of a ready sandbox until it ends. Paused time is not billed, and nothing accrues before the sandbox is handed over. - Plans: Free $0.00/month (Pay as you go at list rates after the $50 starting credit. 20 sandboxes at once, sessions up to 1 hour, sizes up to Large (8 vCPU), 20 GB of data out a month); Pro $50.00/month (Set up by email. 'More compute each month than the plan costs' (amount not stated), up to 100 sandboxes at once, verified snapshot builds, harness adapters, email support); Max $250.00/month (Set up by email. 5x the usage of Pro (or 20x for $500), up to 500 sandboxes at once (1,000 at $500), capacity held warm, shared Slack channel; usage past the allowance below list by commitment) - Free tier: $50 of free credit to start; the owner claims it once in the browser. - Isolation: Isolated sandbox per run; the isolation technology is not stated - Pause and resume: yes - Snapshots: yes - GPU: no - Custom images: yes - Network controls: yes - Max vCPU: 8 - Max memory GiB: 32 - SDKs: JavaScript/TypeScript, Python, CLI, MCP server, REST API - Regions: Not stated on the pages checked; the API reports a location code such as us-east. - Open source: no - Self-hostable: no - Best for: Evaluation and RL runs that fork many identical sandboxes and need environment failures kept apart from agent failures; Teams that want a price quote and a hard spend cap before each batch starts; Running GitHub Actions or GitLab CI jobs in sandboxes from the same account - Watch out for: Free accounts are limited to 1-hour sessions and 20 sandboxes at once; Paid plans are arranged by email and their included usage is not stated in dollars; The isolation technology is not documented; Some accounts must answer setup questions before starting work - Sources: https://www.promigence.ai/pricing https://www.promigence.ai/docs/limits https://www.promigence.ai/security https://www.promigence.ai https://www.promigence.ai/docs/llms-full.txt - Facts checked: 2026-11-13 ## gVisor gVisor is a building block, not a ready-made sandbox service: an application kernel written in Go that runs in userspace and implements a Linux-like interface, so sandboxed containers do not talk to the host kernel directly. Its OCI runtime, runsc, plugs into Docker, containerd and Kubernetes. Many sandbox products and GKE Sandbox use it underneath. It is free Apache-2.0 software for x86_64 and ARM64 Linux. - Page: https://aisandboxhunt.com/gvisor - Company: Google - Website: https://gvisor.dev - Categories: AI agent sandboxes - Score: 54/100, #53 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: Self-host · no list price for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Billing increment: Not applicable: free software run on your own hardware - Billed while: Not applicable: no usage charges - Free tier: Free open-source software under the Apache-2.0 license. - Isolation: Userspace application kernel (gVisor) between the container and the host kernel - Custom images: yes - SDKs: OCI runtime (runsc), Docker, containerd, Kubernetes RuntimeClass - Regions: Runs on your own x86_64 or ARM64 Linux hosts. - Open source: yes (Apache-2.0) - Self-hostable: yes - Best for: Hardening containers that run untrusted code without moving to full VMs; Adding a sandboxed runtime class to an existing Docker or Kubernetes setup; Teams building their own sandbox platform - Watch out for: Building block only: no API, SDK or sandbox lifecycle features on its own; Not a VM; it reduces but does not remove the host kernel attack surface; Linux only, x86_64 and ARM64 - Sources: https://github.com/google/gvisor https://github.com/google/gvisor/releases/tag/release-20261109.0 - Facts checked: 2026-11-13 ## Kubernetes Agent Sandbox Agent Sandbox is a Kubernetes SIG Apps project that adds a Sandbox custom resource and controller for long-running, stateful single-pod workloads with a stable identity and persistent storage. Extensions add templates, claims and pre-warmed pools for fast allocation, plus Go and Python SDKs and an optional router. It orchestrates pods and leaves kernel isolation to a runtime class such as gVisor or Kata Containers. Google offers it on GKE at no extra charge beyond normal GKE resource pricing. - Page: https://aisandboxhunt.com/kubernetes-agent-sandbox - Company: Kubernetes SIG Apps - Website: https://agent-sandbox.sigs.k8s.io - Categories: AI agent sandboxes - Score: 54/100, #54 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: Self-host · no list price for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Billing increment: Not applicable: free software run on your own cluster - Billed while: Not applicable: you pay only for the cluster resources the pods use - Free tier: Free open-source software under Apache-2.0. On GKE it is offered at no extra charge; normal GKE pricing applies to the resources you create. - Isolation: Kubernetes pod per sandbox; stronger isolation comes from a RuntimeClass such as gVisor or Kata Containers (GKE uses gVisor or microVMs) - Pause and resume: yes - GPU: yes - Custom images: yes - SDKs: Go, Python, Kubernetes API (kubectl/YAML) - Regions: Any Kubernetes 1.25+ cluster you run (1.29+ recommended), including GKE. - Open source: yes (Apache-2.0) - Self-hostable: yes - Best for: Teams already running Kubernetes that want a standard API for per-agent sandboxes; Reinforcement learning and evaluation loops that need fast claims from warm pools; GKE users who want managed sandbox support at no extra charge - Watch out for: Plain pods share the host kernel; you must configure gVisor, Kata or another runtime for untrusted code; You operate the cluster, controller upgrades and networking yourself; Requires Kubernetes 1.25 or later with CEL validation for CRDs - Sources: https://github.com/kubernetes-sigs/agent-sandbox https://docs.cloud.google.com/kubernetes-engine/docs/how-to/agent-sandbox https://docs.cloud.google.com/kubernetes-engine/docs/how-to/how-install-agent-sandbox https://github.com/kubernetes-sigs/agent-sandbox/releases/tag/v1.0.11 - Facts checked: 2026-11-13 ## NVIDIA OpenShell OpenShell runs autonomous agents in sandboxes where a declared policy controls file access, system calls and every outbound network connection, and credentials are injected only into requests to approved endpoints. A gateway manages sandboxes on one compute driver: Docker, Podman, Kubernetes (via the Agent Sandbox controller) or an opt-in microVM driver. Sandboxes can request GPUs, run any OCI image, and be stopped and started while keeping their workspace. It is Apache-2.0 and moving to the CNCF. - Page: https://aisandboxhunt.com/nvidia-openshell - Company: NVIDIA - Website: https://github.com/NVIDIA/OpenShell - Categories: AI agent sandboxes - Score: 54/100, #55 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: Self-host · no list price for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Billing increment: Not applicable: free software run on your own infrastructure - Billed while: Not applicable: you pay only for your own hardware or cluster - Free tier: Free open-source software under the Apache-2.0 license. - Isolation: Container per sandbox on Docker, Podman or Kubernetes with Landlock and seccomp policy enforcement; optional microVM driver (KVM or Apple Hypervisor) for a VM boundary - Pause and resume: yes - GPU: yes - Custom images: yes - Network controls: yes - SDKs: Python, TypeScript, Go, Rust, CLI - Regions: Runs where you deploy the gateway: Linux, macOS on Apple Silicon, Windows with WSL 2 (experimental), or a Kubernetes cluster. - Open source: yes (Apache-2.0) - Self-hostable: yes - Best for: Running agents with per-sandbox file, syscall and network policies that need human review to widen; Keeping real API credentials out of the agent's environment; GPU agent workloads on Docker, Podman or Kubernetes - Watch out for: Default drivers use containers; VM isolation needs the opt-in microVM driver; Kubernetes deployments require a network plugin that enforces NetworkPolicy; Anonymous telemetry is on unless disabled - Sources: https://github.com/NVIDIA/OpenShell https://docs.nvidia.com/openshell/latest/how-it-works/sandboxes/runtimes https://docs.nvidia.com/openshell/latest/how-it-works/sandboxes/overview https://github.com/NVIDIA/OpenShell/releases/tag/v0.1.8 - Facts checked: 2026-11-13 ## Riza Code Interpreter API Riza runs short scripts or functions sent over an API in a fresh WebAssembly sandbox per execution, with no filesystem or network access unless you grant it. It supports Python, JavaScript, TypeScript, Ruby and PHP, with third-party packages for Python and JavaScript through custom runtimes. Pricing is by request count, not compute time: Hobby is free for up to 100,000 requests a month and Pro is $250 a month for 1,000,000. Enterprise adds self-hosting. - Page: https://aisandboxhunt.com/riza - Company: Riza - Website: https://riza.io - Categories: AI agent sandboxes - Score: 52/100, #56 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: Self-host · no list price for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Billing increment: per request (each code execution) - Billed while: Each execution request counts against the plan's monthly request quota; execution time and resources are not metered separately. - Plans: Hobby $0.00/month (Up to 100,000 requests and 100 GB data transfer per month; 30-second execution limit per request; 3 custom runtimes); Pro $250.00/month (1,000,000 requests and 500 GB data transfer per month; up to 15-minute execution limit per request; 50 custom runtimes. Higher limits use custom usage-based pricing) - Free tier: Hobby plan is free with up to 100,000 requests per month, 30-second executions and 3 custom runtimes. - Isolation: WebAssembly sandbox per execution; modules are not shared across executions - Pause and resume: no - Snapshots: no - GPU: no - Custom images: yes - Network controls: yes - Max session hours: 0.25 - Max memory GiB: 0.125 - SDKs: Python, Node.js, Go, REST API, MCP server - Regions: Not stated - Open source: no - Self-hostable: yes - Best for: Running short LLM-generated scripts where per-request pricing is simpler than compute metering; Multi-language snippets in Python, JavaScript, TypeScript, Ruby or PHP; Free hobby projects within 100,000 executions a month - Watch out for: 128 MiB memory and 30-second executions on the free plan; No shell, persistent state or general Linux environment; Third-party packages only for Python and JavaScript; Jump from free to $250 per month with no published plan in between - Sources: https://riza.io/pricing https://docs.riza.io/interpreters/limits https://docs.riza.io/reference/security https://docs.riza.io/getting-started/client-libraries - Facts checked: 2026-11-13 ## SandBase SandBase, run by Singapore-based AXIOMCLOUD, sells AI models, data APIs and managed agents on one prepaid balance, and adds an agent sandbox that speaks the E2B protocol. You keep the official E2B JavaScript or Python SDK and point it at SandBase's endpoint with a SandBase key, then create an instance from a template, run commands, read and write files, reach ports inside it, and pause and reconnect where the template supports it. SandBase describes the sandbox as a gateway: the runtime and isolation come from the underlying provider behind each template, which it does not name. The public templates are 2 vCPU with 2 GB of memory. Sandboxes are free for now, with possible fair-use caps. - Page: https://aisandboxhunt.com/sandbase - Company: AXIOMCLOUD PTE. LTD. - Website: https://www.sandbase.ai - Categories: AI agent sandboxes - Score: 49/100, #57 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: No public rate for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Billing increment: Not applicable: sandboxes are currently free - Billed while: Not billed for now; creating, running and stopping sandboxes does not use the account balance - Plans: Start $0.00/month (Starter credits for models and APIs; sandboxes are free for now and fair-use caps may apply. The credit amount is not stated) - Free tier: Sandboxes are free for now, with possible fair-use caps. SandBase says it will announce pricing before this changes. - Isolation: Not stated; SandBase is a gateway and says isolation depends on the underlying provider and template - Pause and resume: yes - SDKs: E2B Python SDK, E2B JavaScript SDK, REST API - Regions: Not stated. - Open source: no - Self-hostable: no - Best for: Trying an E2B-style sandbox at no cost by changing only the key, API URL and domain; Builders who already buy models or data APIs from SandBase and want code execution on the same key; Short-lived agent tasks that run commands and exchange files - Watch out for: The isolation technology and the underlying provider are not disclosed, so you cannot judge the security boundary from SandBase's pages; The free offer is temporary and future sandbox pricing is unknown; Upstream E2B templates may not be available; you must pick from the templates visible to your organization - Sources: https://www.sandbase.ai/pricing https://www.sandbase.ai/sandbox https://www.sandbase.ai/docs/sandbox/quickstart https://raw.githubusercontent.com/sandbaseai/sandbase/main/catalog/sandbox.json https://www.sandbase.ai/terms - Facts checked: 2026-11-15 ## Azure Container Apps dynamic sessions Dynamic sessions in Azure Container Apps hand out prewarmed sandboxes from a session pool, each isolated by a Hyper-V boundary. Built-in code interpreter pools run Python, Node.js or Shell and cost $0.03 per session-hour, billed in one-hour increments from allocation to deallocation. Custom container pools run your own image on Dedicated plan compute instead. Sessions end after an idle cooldown of 5 to 60 minutes. - Page: https://aisandboxhunt.com/azure-dynamic-sessions - Company: Microsoft - Website: https://learn.microsoft.com/azure/container-apps/sessions - Categories: AI agent sandboxes - Score: 48/100, #58 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: No public rate for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Fixed-size price per hour: $0.03 (One built-in code interpreter session (Python, Node.js or Shell); vCPU and memory per session not stated) - Billing increment: 1 hour per allocated session - Billed while: From when a session is allocated until it is deallocated (after the idle cooldown or an explicit delete), rounded up to whole hours. - Free tier: No dynamic sessions free allowance is listed. New Azure customers can get a $200 credit to use within 30 days. - Isolation: Hyper-V isolation per session - GPU: no - Custom images: yes - Network controls: yes - SDKs: REST API, Azure CLI, LangChain, LlamaIndex, AutoGen and Semantic Kernel integrations, MCP server (Python pools) - Regions: 38 Azure regions across the Americas, Europe, Asia Pacific, the Middle East and Africa. - Open source: no - Self-hostable: no - Best for: Running LLM-generated Python or JavaScript inside an existing Azure deployment; Agent frameworks with ready integrations such as LangChain or Semantic Kernel; Many short interactive sessions that need near-instant start from a prewarmed pool - Watch out for: Hourly billing increments make very short sessions relatively expensive; Each code execution stops after 220 seconds; Session vCPU and memory sizes for built-in interpreters are not published; Custom container pools bill whole Dedicated plan instances rather than per session - Sources: https://azure.microsoft.com/en-us/pricing/details/container-apps/ https://learn.microsoft.com/en-us/azure/container-apps/billing https://learn.microsoft.com/en-us/azure/container-apps/sessions-code-interpreter https://learn.microsoft.com/en-us/azure/container-apps/session-pool https://learn.microsoft.com/en-us/azure/container-apps/sessions https://azure.microsoft.com/en-us/pricing/purchase-options/azure-account - Facts checked: 2026-11-13 ## Cloudflare Dynamic Workers Dynamic Workers let a Cloudflare Worker load and run code supplied at runtime in a new V8 isolate, as a lightweight alternative to containers. The parent Worker decides which bindings the code receives and can block, filter or rewrite its outbound requests. Billing uses Workers Standard rates for requests and CPU milliseconds (including isolate startup), plus a charge per unique Dynamic Worker created each day. It runs JavaScript and WebAssembly only, with 128 MB of memory per isolate. - Page: https://aisandboxhunt.com/cloudflare-dynamic-workers - Company: Cloudflare - Website: https://developers.cloudflare.com/dynamic-workers/ - Categories: AI agent sandboxes - Score: 48/100, #59 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $12.00 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.072 (active CPU only) - Price per GiB-hour of memory: $0.00 - Billing increment: per CPU millisecond, plus per request and per unique Dynamic Worker per day - Billed while: CPU time only: isolate startup plus active execution, excluding time waiting on I/O. Wall-clock duration and memory are not billed. - Plans: Workers Paid $5.00/month (Required for Dynamic Workers. Includes per month 30 million CPU milliseconds, 10 million requests and 1,000 unique Dynamic Workers; then $0.02 per million CPU ms, $0.30 per million requests and $0.002 per Dynamic Worker per day) - Free tier: None. Dynamic Workers are only available on the Workers Paid plan ($5/month minimum), which includes a monthly usage allowance. - Isolation: V8 isolate per Dynamic Worker - GPU: no - Custom images: no - Network controls: yes - Max memory GiB: 0.125 - SDKs: JavaScript/TypeScript (Worker Loader binding inside a Worker), WebAssembly - Regions: Runs on Cloudflare's global network; region choice is not described. - Open source: no - Self-hostable: no - Best for: Running LLM-written JavaScript against an API you expose, instead of many tool calls; Very high volumes of short executions where millisecond startup matters; Platforms on Cloudflare that run code uploaded by their users - Watch out for: JavaScript and WebAssembly only; no Python, shell or Linux processes; 128 MB memory per isolate and up to 5 minutes of CPU per invocation; Requires the Workers Paid plan; there is no free use; Loading code without a stable ID bills a new Dynamic Worker on every call - Sources: https://developers.cloudflare.com/dynamic-workers/pricing/ https://developers.cloudflare.com/workers/platform/pricing/ https://developers.cloudflare.com/workers/platform/limits/ https://developers.cloudflare.com/dynamic-workers/ https://developers.cloudflare.com/dynamic-workers/usage/egress-control/ - Facts checked: 2026-11-13 ## Anthropic sandbox-runtime sandbox-runtime (srt) restricts what a process and its children can read, write and reach on the network, using the operating system's own sandboxing: Seatbelt on macOS, bubblewrap on Linux and a dedicated user account with a Windows Filtering Platform fence on Windows. Network traffic goes through local proxies that enforce domain allow and deny lists. It was built for Claude Code and is published as an npm package that works as a CLI or a library. It is a research preview under Apache-2.0. - Page: https://aisandboxhunt.com/anthropic-sandbox-runtime - Company: Anthropic - Website: https://github.com/anthropics/sandbox-runtime - Categories: AI agent sandboxes - Score: 46/100, #60 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: Self-host · no list price for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Billing increment: Not applicable: free software run on your own machine - Billed while: Not applicable: no usage charges - Free tier: Free open-source software under the Apache-2.0 license. - Isolation: OS-level process sandbox (sandbox-exec/Seatbelt on macOS, bubblewrap on Linux, restricted user account on Windows) with proxy-based network filtering - Network controls: yes - SDKs: CLI, TypeScript/JavaScript library - Regions: Runs locally on macOS, Linux or Windows. - Open source: yes (Apache-2.0) - Self-hostable: yes - Best for: Limiting what a local coding agent, MCP server or shell command can read, write and contact; Adding domain allow lists to processes on a developer machine without running containers; Embedding OS-level sandboxing in a TypeScript tool - Watch out for: Shares the host kernel; weaker than VM-based isolation for hostile code; Beta research preview; APIs and configuration may change; Behavior and limits differ between macOS, Linux and Windows - Sources: https://github.com/anthropics/sandbox-runtime https://github.com/anthropics/sandbox-runtime/releases/tag/v0.0.83 - Facts checked: 2026-11-13 ## Judge0 Judge0 is a GPL-3.0 code execution system that compiles and runs submissions in 60+ languages inside a sandbox with configurable time and memory limits. You can self-host it with Docker at no cost or use the hosted cloud API. The cloud is priced per submission, not per hour: a free Basic plan allows 50 submissions a day, daily-quota plans start at $33.75 a month, with $0.0013 for each submission over the daily quota. It suits running single programs, not long-lived environments. - Page: https://aisandboxhunt.com/judge0 - Company: Judge0 d.o.o. - Website: https://judge0.com - Categories: AI agent sandboxes - Score: 46/100, #61 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: Self-host · no list price for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Billing increment: per submission (cloud); self-hosted is free software - Billed while: Cloud: each submission counts against a daily quota or is billed per submission; run time is not metered. Self-hosted: you pay only for your own servers. - Plans: Basic (cloud) $0.00/month (50 submissions per day at no cost; no credit card required); Pro (cloud) $33.75/month (2,000 submissions per day; $0.0013 per extra submission; community and email support); Ultra (cloud) $67.50/month (5,000 submissions per day; $0.0013 per extra submission; community and email support); Mega (cloud) $133.75/month (10,000 submissions per day; $0.0013 per extra submission; community and email support) - Free tier: Self-hosting the open-source version is free with no limits; the hosted Basic plan includes 50 submissions per day at no cost. - Isolation: Each submission runs in the isolate sandbox (Linux namespaces and control groups) inside Judge0's containers - Pause and resume: no - Snapshots: no - GPU: no - Network controls: yes - SDKs: REST API, Python - Regions: Not stated - Open source: yes (GPL-3.0) - Self-hostable: yes - Best for: Coding-assessment, judge and e-learning platforms that run single programs in many languages; Teams that want to self-host a code execution API at no license cost; Low-volume use on the free 50-submissions-per-day cloud plan - Watch out for: Built for one-shot program runs, not persistent or interactive environments; Process-level sandbox rather than a VM boundary; versions up to 1.13.0 had critical security flaws; Last tagged release dates from 2024; GPL-3.0 license applies if you modify and distribute it - Sources: https://judge0.com https://github.com/judge0/judge0 https://github.com/judge0/judge0/blob/master/CHANGELOG.md - Facts checked: 2026-11-14 ## AWS Nitro Sandboxes Nitro Sandboxes is an EC2 feature that runs each workload in its own lightweight VM on the Nitro Hypervisor, next to a parent EC2 instance that lends it vCPUs and memory. On EKS a pod opts in with one line in its spec, and on ECS Managed Instances each task can run in its own sandbox; on plain EC2 you launch them yourself with Firecracker. Confidential mode hides sandbox memory from the parent instance and adds attestation that gates key release in AWS KMS, while direct mode trusts the parent. There is no extra charge: you pay for the EC2 instances (and any EKS Auto Mode or ECS Managed Instances fee). - Page: https://aisandboxhunt.com/aws-nitro-sandboxes - Company: Amazon Web Services - Website: https://aws.amazon.com/about-aws/whats-new/2026/11/nitro-sandboxes-eks-ecs-ec2/ - Categories: AI agent sandboxes - Score: 45/100, #62 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: No public rate for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Billing increment: Not applicable: no charge for the sandboxes themselves; EC2 instances are billed at standard rates - Billed while: Not applicable: you pay for the parent EC2 instances while they run, however many sandboxes they host - Free tier: The sandbox feature and its EKS and ECS integrations cost nothing extra, but there is no free compute: you pay standard EC2 prices for the parent instances. - Isolation: Lightweight VM per sandbox on the Nitro Hypervisor, launched through Firecracker on a parent EC2 instance; optional confidential mode with attestation - Pause and resume: no - Snapshots: no - GPU: no - Custom images: yes - Network controls: yes - SDKs: Kubernetes API (kubectl/YAML) on EKS, Amazon ECS API, Firecracker on EC2 - Regions: All commercial AWS Regions, on most Nitro-based EC2 instance types (not flex instances). - Open source: no - Self-hostable: no - Best for: Teams on EKS or ECS that want a VM boundary per agent session or pod without a dedicated instance each; Running untrusted pull request code or LLM-generated code next to other tenants on shared instances; Confidential workloads that must stay hidden from the instance operator, with attestation-gated keys - Watch out for: Not a hosted API: you run and pay for the EC2 instances, clusters and orchestration yourself; No pause and resume, snapshot and restore, or live migration; No GPU passthrough; Without a block rule on the parent, a sandbox whose traffic is routed through the parent can reach the parent's instance metadata and its IAM role credentials - Sources: https://aws.amazon.com/about-aws/whats-new/2026/11/nitro-sandboxes-eks-ecs-ec2/ https://aws.amazon.com/blogs/compute/nitro-sandboxes-hardware-isolation-for-every-eks-pod-ecs-task-and-ec2-workload https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/amazon-ec2-nitro-sandboxes.html https://docs.aws.amazon.com/eks/latest/userguide/sandboxing.html - Facts checked: 2026-11-15 ## Beam Beam sandboxes are disposable containers for agent code, controlled from Python or TypeScript SDKs, with process, filesystem, networking and snapshot APIs; a sandbox can also run in a microVM, and GPUs can be attached. CPU and memory bill by the millisecond while a sandbox runs, at $0.05 per vCPU-hour and $0.01 per GiB-hour on the default Performance tier, with cheaper Shared and Spot tiers. Beam also sells persistent Linux VMs for agents at the same rates, GPU machines and serverless endpoints. Plans start at $20 a month with $20 of credit, and the engine behind Beam is open source (Beta9) and can be self-hosted. - Page: https://aisandboxhunt.com/beam - Company: Beam - Website: https://beam.cloud - Categories: AI agent sandboxes, GPU sandboxes, Dev boxes for AI agents - Score: 45/100, #63 of 63 AI agent sandboxes (method 1.3 (2026-11-14)) - List price per month: $24.64 for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30% - Price per vCPU-hour: $0.05 - Price per GiB-hour of memory: $0.01 - GPU prices per hour: NVIDIA RTX 4090 24 GB (GPU only; CPU and RAM extra) $0.69, NVIDIA RTX 5090 32 GB (GPU only; CPU and RAM extra; committed spend) $1.09 - Storage per GiB-month: $0.00 - Billing increment: per millisecond - Billed while: Only while the sandbox is running, for its requested CPU and memory, until it is terminated or its TTL (default 600 seconds) expires. - Plans: Developer $20.00/month ($20 in credits each month. 30 concurrent vCPUs and 5 concurrent GPUs across VMs, sandboxes and serverless apps; 1 seat; 30-day log retention); Team $89.00/month ($89 a month plus usage. 1,000 concurrent vCPUs, 50 concurrent GPUs, 3 seats included ($25 per extra seat), private Slack channel) - Free tier: No free tier or trial credit is stated on the pricing page; the cheapest plan is $20 a month with $20 of credit. - Isolation: Isolated container by default; a sandbox can run in a microVM with its own kernel (use_vm=True). Persistent VMs are microVMs. - Snapshots: yes - GPU: yes - Custom images: yes - Network controls: yes - SDKs: Python, JavaScript/TypeScript, CLI, MCP server - Regions: Not stated for sandboxes. Bring-your-own-cloud runs in your AWS, GCP or Azure account. - Open source: yes (AGPL-3.0) - Self-hostable: yes - Best for: Agent sandboxes that sometimes need a GPU attached; Cost-sensitive batch work that can use the Spot or Shared CPU tiers; Teams that may later self-host the open-source engine - Watch out for: Sandboxes default to containers; microVM isolation must be requested; No free tier; the entry plan is $20 a month; The Developer plan caps concurrency at 30 vCPUs and 5 GPUs; Storage is free today but planned to be billed - Sources: https://www.beam.cloud/pricing https://www.beam.cloud/pricing.md https://docs.beam.cloud/v2/sandbox/overview https://github.com/beam-cloud/beta9 - Facts checked: 2026-11-13 # Dev boxes for AI agents Persistent machines that an agent or developer keeps for days: the disk survives stops, and most can pause when idle. Includes hosted cloud dev environments. Ranking: https://aisandboxhunt.com/dev-boxes ## Sailboxes Sailboxes are isolated Linux VMs from Sail Research, the inference provider, built for agents that run for hours or days. There is no runtime limit: a Sailbox can sleep or pause with memory and running processes saved, and checkpoints can start new Sailboxes from the same state. Billing follows actual use rather than reserved size: CPU time, physical RAM held and disk used are metered while the Sailbox runs, up to 64 vCPUs and 256 GiB per box. The free plan gives $5 of credit a month once a payment method is attached; Pro costs $250 a month with $100 of included credit. - Page: https://aisandboxhunt.com/sailboxes - Company: Sail Research - Website: https://sailresearch.com/sailboxes - Categories: Dev boxes for AI agents - Score: 86/100, #1 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $8.36 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Price per vCPU-hour: $0.015 (active CPU only) - Price per GiB-hour of memory: $0.008 - Storage per GiB-month: $0.30 - Billing increment: Sampled every second and reported in roughly one-minute batches; not rounded up to whole minutes - Billed while: CPU time, RAM held and disk used are billed only while the Sailbox runs; sleeping, paused, checkpointing and cold-starting time is not billed. Volume storage is billed for every hour a volume exists. - Plans: Free $0.00/month ($5 of credit every month once a payment method is attached; usage paid from prepaid credits; up to 4 seats; 200 reserved vCPUs across active Sailboxes; $0.01 fee per Sailbox created); Pro $250.00/month ($100 of included credit every month plus an extra $150 in the first month; 10,000 reserved vCPUs across active Sailboxes; creation fees waived; unlimited seats; priority support) - Free tier: Free plan: $5 of credit every month after a payment method is attached; usage beyond it is paid from prepaid credits. - Isolation: Separate Linux VM per Sailbox, each with its own kernel - Pause and resume: yes - Snapshots: yes - Custom images: yes - Network controls: yes - Max vCPU: 64 - Max memory GiB: 256 - SDKs: Python, TypeScript, Rust, CLI, REST API - Regions: Not stated - Open source: no - Self-hostable: no - Best for: Agents and jobs that run for days without a runtime cap; Workloads with spiky or unknown memory needs, since only RAM actually held is billed; Long tasks that need to pause and later resume with processes and memory intact - Watch out for: Free plan credit needs a payment method attached, and each new Sailbox costs $0.01 on Free; Billed memory includes the Linux file cache inside the Sailbox, so an idle running box still accrues charges; Regions and GPU support are not documented on the pages checked; Pro costs $250 a month, a large step up from the free plan - Sources: https://docs.sailresearch.com/pricing https://docs.sailresearch.com/sailboxes-pricing.md https://sailresearch.com/sailboxes - Facts checked: 2026-11-13 ## boxd boxd runs persistent Linux machines in microVMs on its own KVM-based VMM, hosted in Frankfurt and New York. Machines default to 2 vCPU, 8 GiB and a 100 GiB disk, can be forked while running, checkpointed and snapshotted, and hibernate after 4 hours without network traffic, keeping their processes. Billing is credit-based and metered per second on active vCPU, resident memory and disk written, not on provisioned size. New organizations get 30 in free credits when they add a payment method, and the platform can also be self-hosted, free up to 16 vCPUs. - Page: https://aisandboxhunt.com/boxd - Company: boxd - Website: https://boxd.sh - Categories: Dev boxes for AI agents - Score: 77/100, #4 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $16.46 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Price per vCPU-hour: $0.047 (active CPU only) - Price per GiB-hour of memory: $0.014 - Storage per GiB-month: $0.07008 - Billing increment: Metered every second, settled every 15 minutes, no minimum charge per machine - Billed while: vCPU only while running and only for active vCPU; RAM on resident memory while running or in standby; disk on space written in every state, including hibernation. - Plans: Pay as you go $0.00/month (Credit-based, no monthly or seat fee. Up to 100 machines per organization, machines up to 8 vCPU / 32 GiB. Auto top-up adds 20 at a time once credits run out (can be changed or turned off)) - Free tier: 30 of free credits (EUR or USD, by the organization's currency) when a payment method is added; nothing is charged when the card is added. - Isolation: Dedicated microVM per machine on boxd's own Rust VMM built on KVM - Pause and resume: yes - Snapshots: yes - Network controls: yes - Max vCPU: 8 - Max memory GiB: 32 - SDKs: Python, TypeScript, CLI, REST API, MCP server - Regions: Frankfurt (Germany) and New York; machines and data in Frankfurt stay in the EU. - Open source: no - Self-hostable: yes - Best for: Dev boxes and agent machines that sit idle much of the day, since idle vCPU is not billed and hibernated machines pay only for disk; Forking a running machine, with memory and processes, to try several approaches in parallel; Teams that want EU hosting or the option to self-host the same platform - Watch out for: Machines larger than 4 vCPU / 16 GiB may need a request by email; Only two regions, Frankfurt and New York; A card is required before the first machine, and auto top-up charges it once free credits run out unless turned off; No GPU support is described - Sources: https://boxd.sh/llms.txt https://boxd.sh/pricing/ https://boxd.sh/ https://docs.boxd.sh/ - Facts checked: 2026-11-13 ## InstaBox InstaBox gives each agent a Firecracker microVM running Debian with Docker, coding agents, a desktop and Chromium in the default template, plus a 20 GiB persistent disk. Machines can be paused with memory kept, stopped, snapshotted, forked and resized, and on paid plans they have no maximum run time by default. vCPU and memory are billed per second only while running; paid plans charge a monthly fee that comes back as usage credit. The free plan runs one 4 vCPU / 8 GiB machine at no charge. - Page: https://aisandboxhunt.com/instabox - Company: InsForge - Website: https://instabox.sh - Categories: Dev boxes for AI agents, AI agent sandboxes - Score: 77/100, #5 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $20.00 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Price per vCPU-hour: $0.004 - Price per GiB-hour of memory: $0.00225 - GPU prices per hour: NVIDIA L4 (paid plans; plus the machine's vCPU and memory) $0.64 - Storage per GiB-month: $0.08 - Billing increment: per second - Billed while: vCPU and memory only while the machine is running (not starting, paused or stopped). On paid plans, persistent disk beyond the plan's included amount is billed while the machine exists, running or not. - Plans: Free $0.00/month (1 machine (any state) at 4 vCPU / 8 GiB, never billed. 20 GiB fixed disk, 5 snapshots, 200 GiB stored data, 100 GiB egress a month. Pauses after 30 minutes idle; each run starts with a 120-minute limit that grows with use up to 480 minutes. Limited templates; no GPUs); Pro $20.00/month ($20/month returned as usage credit. 100 running machines (500 total), sizes 2/4, 4/8 or 8/16 vCPU/GiB, disk up to 200 GiB per machine, 200 GiB disk included across machines, 1 TiB egress a month, email support); Builder $100.00/month ($100/month returned as usage credit. 200 running machines (1,000 total), up to 16 vCPU / 32 GiB, disk up to 500 GiB per machine, 1 TiB disk included, 5 TiB egress a month); Team $500.00/month ($500/month returned as usage credit. 500 running machines, unlimited total, up to 32 vCPU / 64 GiB, disk up to 1 TiB per machine, 5 TiB disk included, 50 TiB egress a month, priority support and SLAs) - Free tier: Free plan: one 4 vCPU / 8 GiB machine that is never billed, with a fixed 20 GiB disk, 5 snapshots and 100 GiB egress a month; it pauses after 30 idle minutes, and each run is limited to 120 minutes at first, growing with use up to 480 minutes. - Isolation: Firecracker microVM per machine - Pause and resume: yes - Snapshots: yes - GPU: yes - Custom images: yes - Max vCPU: 32 - Max memory GiB: 64 - SDKs: TypeScript, Python, CLI, REST API, MCP server - Regions: Not stated; the API has no region setting. - Open source: no - Self-hostable: no - Best for: Persistent per-agent Linux machines with Docker, coding agents and a desktop already installed; Low per-hour compute cost, with small plan fees returned as usage credit; Android app testing with adb and a visible phone screen, on every plan - Watch out for: Free plan is one machine, paused after 30 idle minutes, with runs limited to 120 minutes at first and at most 480; Paid plans start at $20/month; there is no fee-free pay-as-you-go option; Disk beyond the plan's included amount is billed even while machines are stopped; Region is not stated and cannot be chosen; outbound bandwidth is capped at about 50 Mbit/s per machine - Sources: https://instabox.sh/pricing https://instabox.sh/docs/pricing.md https://instabox.sh/docs/billing.md https://instabox.sh/docs/limits.md https://instabox.sh/llms.txt https://instabox.sh/llms-full.txt - Facts checked: 2026-11-14 ## MIOSA MIOSA offers isolated sandboxes for code, persistent desktop computers for browser and computer-use agents, and a path to publish the result as a deployment with databases and custom domains. Sandboxes run in Firecracker microVMs by default, keep their disk when stopped, and can pause, snapshot and fork. Sandboxes and desktops are billed a flat price per size, per second while running, from $0.0316 an hour for 1 vCPU / 0.5 GB to $0.7141 for 16 vCPU / 32 GB. Every plan, including the free one, credits its monthly price back as usage. - Page: https://aisandboxhunt.com/miosa - Company: MIOSA - Website: https://miosa.ai - Categories: Dev boxes for AI agents, AI agent sandboxes, Browser sandboxes for AI agents - Score: 77/100, #6 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $20.00 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Fixed-size price per hour: $0.0893 (small: 2 vCPU, 4 GB) - Price per browser-hour: $0.0446 - Storage per GiB-month: $0.09 - Billing increment: Per second - Billed while: Flat price of the size while the sandbox or desktop runs; stopped or paused machines pay no compute, only disk. - Plans: Free $0.00/month ($1 of usage every month. 5 sandboxes and 1 computer at once, machines up to 4 vCPU / 8 GB, 8 GB and 4 vCPU running in total, 5 GB egress a month, 1 seat. Machines pause at a $0 balance; credits can be added from $5); Developer $20.00/month ($20 a month counted toward usage. 50 sandboxes and 5 computers at once, machines up to 4 vCPU / 8 GB, 128 GB and 64 vCPU running in total, 100 GB egress a month, 1 seat); Business $500.00/month ($500 a month counted toward usage. 250 sandboxes and 50 computers at once, machines up to 8 vCPU / 16 GB, 512 GB and 256 vCPU running in total, 1 TB egress a month, 10 seats, dedicated vCPU sizes); Enterprise $2,000.00/month ($2,000 a month counted toward usage. 1,000 sandboxes and 200 computers at once, machines up to 16 vCPU / 32 GB, 1,024 GB and 512 vCPU running in total, 5 TB egress a month, unlimited seats) - Free tier: Free plan: $1 of usage every month with no overage; machines pause at a $0 balance. - Isolation: Firecracker microVM by default; an alternative SOMA runtime is also offered - Pause and resume: yes - Snapshots: yes - GPU: no - Network controls: yes - Max vCPU: 16 - Max memory GiB: 32 - SDKs: TypeScript, Python, Go, Rust, Java, Elixir, CLI, MCP server, REST API - Regions: US West (Los Angeles), US East (New York) and US East 2 (Miami). Bring-your-own-cloud on AWS or Google Cloud is set up with MIOSA. - Open source: no - Self-hostable: no - Best for: Agent products that need code sandboxes, desktops and production deployments under one account; Small, cheap sandboxes for short code runs, from about $0.03 an hour; Computer-use agents that need a desktop priced the same as a sandbox - Watch out for: Sizes above 4 vCPU / 8 GB need the $500 or $2,000 monthly plan; Egress is billed from the first GB; Only US regions; No GPU sandboxes yet - Sources: https://miosa.ai/pricing https://miosa.ai/docs/develop/sandboxes.md https://miosa.ai/docs/regions.md https://miosa.ai/docs/develop/sandboxes/timeouts.md https://miosa.ai/docs/known-limitations/ - Facts checked: 2026-11-13 ## Archil Archil sells durable disks for AI workloads and runs compute next to them: persistent sandboxes and one-shot serverless tool calls. Each sandbox is a Firecracker microVM whose root filesystem sits on an Archil disk, so installed packages survive stop, start and pause, and sandboxes can be forked. Sessions pause after at most 24 hours with memory intact and can be resumed. Sandbox CPU and memory are billed at $0.055 per vCPU-hour and $0.020 per GiB-hour on what is allocated, with paid plans set as monthly minimums that usage counts toward. - Page: https://aisandboxhunt.com/archil - Company: Archil - Website: https://archil.com - Categories: Dev boxes for AI agents, AI agent sandboxes - Score: 73/100, #7 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $41.80 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Price per vCPU-hour: $0.055 - Price per GiB-hour of memory: $0.02 - Storage per GiB-month: $0.20 - Billing increment: Not stated - Billed while: CPU and memory on what a running sandbox is allocated; stopped and paused sandboxes accrue no compute charges. Disk storage is billed separately for as long as data is kept. - Plans: Free $0.00/month (Monthly allowance: 1 vCPU-hour and 2 GiB-hours of sandbox compute, 1,000 serverless tool calls, 5 disks of up to 10 GB each. Up to 5 concurrent sandboxes, sessions of at most 30 minutes, egress limited to a managed allowlist. Hard caps, no overage); Developer $10.00/month ($10 monthly minimum that metered usage counts toward. 1,000 file systems, 20 concurrent sandboxes, external mounts, your own egress firewall rules and credential brokering, private sandbox images); Team $500.00/month ($500 monthly minimum that metered usage counts toward. 25,000 file systems, 250 concurrent sandboxes (500 more for $500 a month), 3 workspaces, SOC 2, shared Slack channel) - Free tier: Free plan: each month 1 vCPU-hour and 2 GiB-hours of sandbox compute (an hour of the default 1 vCPU / 2 GiB sandbox), 1,000 tool calls and 5 disks of up to 10 GB; hard caps with no overage. - Isolation: One Firecracker microVM per sandbox with its own Linux kernel (KVM) - Pause and resume: yes - Snapshots: yes - Custom images: yes - Network controls: yes - Max session hours: 24 - Max vCPU: 32 - Max memory GiB: 64 - SDKs: TypeScript, Python, REST API, CLI - Regions: Disks in AWS us-east-1, us-west-2 and eu-west-1 and GCP us-central1; sandboxes run in Archil's public AWS regions. Other regions and bring-your-own-cloud on request or on Enterprise. - Open source: no - Self-hostable: no - Best for: Agent workspaces whose files and installed packages must outlive any single VM; Sharing one data set between sandboxes, your own servers and S3 tools; Agent bash tools billed per call rather than per machine-hour - Watch out for: Compute is billed on allocated vCPU and memory while a sandbox runs, not on actual use; Sessions pause after 24 hours and must be resumed; Free sessions last at most 30 minutes; Free sandboxes can reach only a managed allowlist of hosts; Paid plans carry a monthly minimum of $10 or $500 - Sources: https://archil.com/pricing https://docs.archil.com/compute/sandboxes/introduction - Facts checked: 2026-11-13 ## boat boat runs each sandbox as its own Ubuntu virtual machine with root, SSH, Docker, a browser desktop and pre-installed coding agents, driven by the boat CLI, Python and TypeScript SDKs or a REST API. Sandboxes come in four fixed sizes billed per second while running, from $0.018 an hour for 2 vCPU / 4 GB to $0.200 for 16 vCPU / 32 GB. A stopped sandbox has no compute charge and keeps its disk for at least 14 days; memory and processes are not kept after a short window. Plans start at $20 a month, and the plan price comes back as VM time. - Page: https://aisandboxhunt.com/boat - Company: boat - Website: https://boat.dev - Categories: Dev boxes for AI agents, AI agent sandboxes - Score: 73/100, #8 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $20.00 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Fixed-size price per hour: $0.018 (small: 2 vCPU, 4 GB RAM, 51 GB disk. Other sizes: default 4 vCPU / 8 GB / 51 GB at $0.036/h, large 8 vCPU / 16 GB / 111 GB at $0.072/h, xlarge 16 vCPU / 32 GB / 251 GB at $0.200/h ($100 plan or higher).) - Billing increment: per second - Billed while: Wall-clock while the sandbox runs, at the size's hourly price. Nothing for compute while stopped. - Plans: $20/mo $20.00/month ($20 of VM time per month (555 hours of the default size), not carried over. 100 sandboxes at once; 12 starts per minute, 60 per hour, 200 per day included; 10 saved snapshots. Sizes up to large. Overage paid from a prepaid wallet); $100/mo $100.00/month ($100 of VM time per month (2,777 hours of the default size). 300 sandboxes at once; 30 / 210 / 840 starts per minute / hour / day included; 50 saved snapshots. Unlocks xlarge); $500/mo $500.00/month ($500 of VM time per month (13,888 hours of the default size). 1,000 sandboxes at once; 65 / 420 / 1,680 starts per minute / hour / day included; 250 saved snapshots); $2000/mo $2,000.00/month ($2,000 of VM time per month (55,555 hours of the default size). 2,000 sandboxes at once; 90 / 600 / 2,400 starts per minute / hour / day included; 1,000 saved snapshots) - Free tier: 7-day free trial on the first plan, card required: 25 compute hours in total, 2 sandboxes at once, small and default sizes only, auto-stop of 2 hours at most, no saved snapshots. There is no free tier without a card. - Isolation: Virtual machine per sandbox with its own kernel, disk and IP address - Pause and resume: yes - Snapshots: yes - GPU: no - Network controls: no - Max vCPU: 16 - Max memory GiB: 32 - SDKs: Python, TypeScript, CLI, REST API - Regions: Germany, Finland and France; no US or Asia region. Snapshot copies can be stored outside the EU. - Open source: no - Self-hostable: no - Best for: Full Linux dev machines for agents, with root, Docker, systemd, SSH and a desktop; Steady, CPU-heavy workloads where a flat hourly price per size is easier to predict than active-CPU billing; Teams in Europe that want sandboxes hosted in Germany, Finland and France - Watch out for: A $20 monthly plan is the minimum; the trial needs a card; No US or Asia region and no GPU on any size; Stopping keeps only the disk; memory and running processes are lost after about 10 minutes; Starts above the plan's daily allowance cost extra and can fail; nothing is queued - Benchmark (aisandboxbench, 2026-10-23): Cold start (median) 1.125 s, rank 6 of 9 - Benchmark (aisandboxbench, 2026-10-23): Agent tool-call replay (median) 43.256 s, rank 9 of 9 - Benchmark (StarSling HPC sandbox benchmarks, 2026-09-29): Node.js web tooling (runs per second) 26.69 runs/s, rank 1 of 12 - Benchmark (StarSling HPC sandbox benchmarks, 2026-09-29): Better Auth cold dependency install 18.7 s, rank 7 of 11 - Sources: https://docs.boat.dev/llms.txt https://docs.boat.dev/pricing https://docs.boat.dev/free-trial https://docs.boat.dev/limits https://docs.boat.dev/security https://boat.dev - Facts checked: 2026-11-14 ## Fly.io Sprites Sprites are hardware-isolated Linux VMs from Fly.io with a persistent 100 GB ext4 disk, checkpoints, and their own HTTPS URL. They pause automatically about 30 seconds after activity stops and wake on the next request. Billing is metered on cumulative CPU time ($0.0385/CPU-hour) and memory actually in use ($0.021875/GB-hour) only while a Sprite is running, plus hot storage while awake and cold storage for as long as data is kept. Optional monthly plans bundle CPU-hours, RAM GB-hours and storage and raise concurrency limits. - Page: https://aisandboxhunt.com/fly-sprites - Company: Fly.io - Website: https://sprites.dev - Categories: Dev boxes for AI agents, AI agent sandboxes - Score: 70/100, #10 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $20.00 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Price per vCPU-hour: $0.0385 (active CPU only) - Price per GiB-hour of memory: $0.02187 - Storage per GiB-month: $0.02 - Billing increment: metered usage (CPU time from cpu.stat, memory in use), rates quoted per hour - Billed while: Only while the Sprite is running (active); CPU on actual CPU time used and memory on actual memory used. Warm and cold (paused) Sprites pay only storage. - Plans: Pay-as-you-go $0.00/month (Usage at standard rates; 10 Sprite creations per minute; community support. Concurrency limit for this tier not stated on the pricing page); Adventurer $20.00/month (Includes 450 CPU-hours, 1,800 RAM GB-hours, 50 GB-months storage; 20 concurrent active and 20 warm Sprites; 60 creations/minute; community support); Veteran $50.00/month (Includes 800 CPU-hours, 3,200 RAM GB-hours, 100 GB-months storage; 50 concurrent active and 50 warm Sprites; community support); Hero $100.00/month (Includes 1,200 CPU-hours, 4,800 RAM GB-hours, 150 GB storage; 100 concurrently running and 100 warm Sprites, cold unlimited; standard email support); Champion $200.00/month (Includes 1,800 CPU-hours, 7,200 RAM GB-hours, 225 GB-months storage; 200 concurrent active and 200 warm Sprites; standard email support); Legend $500.00/month (Includes 3,200 CPU-hours, 12,500 RAM GB-hours, 400 GB-months storage; 500 concurrent active and 500 warm Sprites; standard email support); Epic $1,000.00/month (Includes 4,800 CPU-hours, 18,750 RAM GB-hours, 600 GB-months storage; 1,000 concurrent active and 1,000 warm Sprites; premium email support); Mythic $2,000.00/month (Includes 7,200 CPU-hours, 28,000 RAM GB-hours, 900 GB-months storage; 2,000 concurrent active and 2,000 warm Sprites; 240 creations/minute; premium email support; arranged via billing@fly.io) - Free tier: Every new Fly.io account gets a one-time $1 trial credit (no credit card, lasts 7 days); invoices under $5 are waived. - Isolation: Hardware-isolated microVM per Sprite - Pause and resume: yes - Snapshots: yes - GPU: no - Network controls: yes - Max vCPU: 8 - SDKs: JavaScript/TypeScript, Go, Python, Elixir, CLI, REST API, Remote MCP server - Regions: Runs on Fly.io infrastructure; the pages checked do not document choosing a region for a Sprite. - Open source: no - Self-hostable: no - Best for: Agents and dev environments that sit idle most of the time, since paused Sprites pay only storage; Workloads with low average CPU use, because CPU is billed on CPU time actually consumed; Persistent per-user or per-agent Linux boxes with checkpoints and their own HTTPS URL - Watch out for: Fixed 8 vCPU and 100 GB disk per Sprite; memory has no published fixed size; No GPU support is described; Code running inside a Sprite does not keep it awake by itself; it pauses after about 30 seconds without requests, output, open TCP connections or an active task; Plan allowances are unit bundles, and plan fees are not covered by account credits - Sources: https://sprites.dev https://fly.io/pricing/ https://fly.io/changelog/new-sprites-plans/ https://fly.io/sprites/release-notes/ https://docs.fly.io/sprites/concepts/lifecycle.md - Facts checked: 2026-11-14 ## Workpaper Workpaper starts a Linux virtual machine with its own kernel from a CLI or JavaScript and Python packages, with Claude Code, Codex and a headless browser installed. There are two sizes: small (2 vCPU, 4 GB, 10 GB disk) and large (4 vCPU, 8 GB, 20 GB disk with Docker) at twice the price. Stopping saves disk and memory, including running programs, and a computer can be forked into an independent copy. Computers pause after an hour idle, and the whole service runs on its own small pool of hardware in Delhi. - Page: https://aisandboxhunt.com/workpaper - Company: Workpaper (individual developer, India) - Website: https://workpaper.co - Categories: Dev boxes for AI agents, AI agent sandboxes - Score: 69/100, #12 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $15.00 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Fixed-size price per hour: $0.01 (Small: 2 vCPU, 4 GB RAM, 10 GB disk (Pro rate). Large (4 vCPU, 8 GB, 20 GB disk, Docker) is $0.020/hour, two small units.) - Billing increment: per second - Billed while: Only while a computer is running. Stopped computers cost nothing, including their saved disk and memory. - Plans: Pro $15.00/month ($15/month, all returned as usage credit: 1,500 small-computer hours or 750 large-computer hours a month. Unused monthly credit expires at renewal. Extra usage packs are $15 for 1,500 hours and never expire. Up to 4 running computers per account. New Pro sign-ups go to a waitlist when the service is at capacity) - Free tier: Free trial with no account and no card: 60 minutes on one small computer, with internet limited to code and package sites and AI model APIs; one trial per device every 30 days. New accounts also get one free hour of credit with full internet. - Isolation: Virtual machine with its own Linux kernel per computer (hypervisor not named) - Pause and resume: yes - GPU: no - Custom images: no - Max vCPU: 4 - Max memory GiB: 8 - SDKs: CLI, JavaScript/TypeScript, Python - Regions: India only (Delhi). The docs expect about 150 ms per keystroke from Europe and 250 ms from the US. - Open source: no - Self-hostable: no - Best for: Cheap short agent sessions on a small VM with coding agents already installed; Users in or near India, where latency is lowest; Trying several fixes in parallel by forking a running computer, memory included - Watch out for: Runs only in India on a small, self-owned hardware pool; you may wait in line for a free computer; No backups and no uptime guarantee; the docs advise against data you cannot afford to lose; No inbound ports, no GPUs and no custom images; Operated by an individual developer; new Pro sign-ups can be waitlisted when at capacity - Sources: https://docs.workpaper.co/pricing https://workpaper.co https://docs.workpaper.co/limits https://docs.workpaper.co/how-it-works https://docs.workpaper.co/free-trial https://workpaper.co/terms - Facts checked: 2026-11-14 ## Sandbox0 Sandbox0 gives each user's agent a persistent cloud computer in a gVisor sandbox whose entire writable root filesystem is kept across pauses and grows automatically as it fills. Compute pauses between tasks and the same sandbox resumes later; by default processes restart fresh, with memory checkpoints available as an experimental option. Compute is billed by the second at $0.015 per GiB-hour of the configured memory limit, with no separate CPU charge, and retained changes cost $0.02 per GiB-month. Network traffic is currently free. - Page: https://aisandboxhunt.com/sandbox0 - Company: Sandbox0 - Website: https://sandbox0.ai - Categories: Dev boxes for AI agents - Score: 67/100, #14 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $13.20 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Price per vCPU-hour: $0.00 - Price per GiB-hour of memory: $0.015 - Storage per GiB-month: $0.02 - Billing increment: Per second (about $0.000004167 per GiB-second), no one-hour minimum - Billed while: Compute while the sandbox runs, including idle running time, on its configured memory limit. Paused sandboxes pay only retained storage. - Plans: Pay as you go $0.00/month (Prepaid funds, no plan fee. Default quotas: 20 running sandboxes per team, unlimited paused sandboxes, 5 sandbox claims per second, 100 API requests per second; increases on request) - Free tier: One-time $5 usage credit for eligible new users after a credit card is saved; adding the card charges nothing. - Isolation: gVisor (runsc) sandbox per user - Pause and resume: yes - Snapshots: yes - Custom images: yes - Network controls: yes - SDKs: Python, JavaScript/TypeScript, Go, CLI, REST API - Regions: Not stated; quotas and templates are shown per region in the dashboard. - Open source: no - Self-hostable: no - Best for: Per-user agent computers where installed tools and system configuration must persist; Bursty workloads that pause between tasks, since paused sandboxes pay only for stored changes; CPU-heavy work at a small memory size, since only memory is metered - Watch out for: gVisor is a user-space kernel, a weaker boundary than a hardware VM; Running processes restart after a default pause; memory checkpoints are experimental; Maximum CPU and memory per sandbox and regions are not published; Egress is free only for now - Sources: https://sandbox0.ai/pricing https://sandbox0.ai/ https://sandbox0.ai/docs/sandbox.md https://sandbox0.ai/llms.txt - Facts checked: 2026-11-13 ## Google Cloud Workstations Cloud Workstations runs each developer's environment as a container on a Compute Engine VM created in your own Google Cloud project and VPC, with a persistent disk kept as the home directory between sessions. You pay Compute Engine rates for the VM and disk, a management fee of $0.05 per vCPU-hour while a workstation runs, and $0.20 an hour per cluster control plane. An e2-standard-2 workstation (2 vCPU, 8 GiB) costs $0.16701142 an hour in us-central1 including the management fee. GPU machine types and custom container images are supported. - Page: https://aisandboxhunt.com/google-cloud-workstations - Company: Google Cloud - Website: https://cloud.google.com/workstations - Categories: Dev boxes for AI agents - Score: 66/100, #15 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $36.74 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Fixed-size price per hour: $0.167 (e2-standard-2: 2 vCPU, 8 GiB (us-central1, VM plus management fee)) - Storage per GiB-month: $0.10 - Billing increment: Compute Engine billing; management fee metered from start until stop - Billed while: VM and management fee while the workstation runs; persistent disk while it exists; the cluster control plane every hour whether workstations are in use or not. - Free tier: New Google Cloud customers get $300 in free credit; there is no free allowance specific to Cloud Workstations. - Isolation: Compute Engine VM per running workstation, with the workstation as a container on it - Pause and resume: yes - Snapshots: yes - GPU: yes - Custom images: yes - Network controls: yes - Max vCPU: 96 - Max memory GiB: 768 - SDKs: gcloud CLI, REST API, Console - Regions: Around 30 Google Cloud regions across the Americas, Europe, the Middle East and Asia-Pacific; prices vary by region. - Open source: no - Self-hostable: no - Best for: Companies that need dev environments inside their own Google Cloud project and VPC; Teams that want admin-controlled custom images and policies for every developer; GPU-backed development environments - Watch out for: The $0.20 an hour cluster fee makes it costly for one or two users; Setup requires a Google Cloud project, cluster and workstation configuration; Processes do not survive a stop: only the home disk persists; Prices vary by region and machine family - Sources: https://cloud.google.com/workstations/pricing https://cloud.google.com/workstations/docs/architecture https://cloud.google.com/free - Facts checked: 2026-11-13 ## Morph Cloud Morph Cloud runs VMs (MorphVM instances and Devboxes) on its Infinibranch runtime, which snapshots a running VM and starts new instances from it, including many parallel branches. Usage is billed in Morph Compute Units (MCUs) at $0.05 each; a running VM uses the largest of its vCPU count, RAM/4 GiB and disk/16 GiB in MCUs per hour. Paused instances stop using compute and pay only for the snapshot. Paid tiers are monthly MCU bundles at a discount. - Page: https://aisandboxhunt.com/morph - Company: Morph Labs (not related to Morph LLM / morphllm.com, which is AutoInfra, Inc.) - Website: https://cloud.morph.so - Categories: Dev boxes for AI agents, AI agent sandboxes - Score: 66/100, #16 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $22.00 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Fixed-size price per hour: $0.05 (1 MCU: 1 vCPU + 4 GB RAM + 16 GB disk; a VM costs max(vCPUs, ceil(RAM/4 GiB), ceil(disk/16 GiB)) MCUs per hour) - Storage per GiB-month: $0.0073 - Billing increment: per second (rates published per MCU-hour) - Billed while: wall-clock while the instance is running; paused instances bill only snapshot storage - Plans: Developer (Free) $0.00/month (0 starting MCUs on the pricing page (buy credits to use compute); MorphVM quota up to 64 vCPU, 256 GB RAM, 1024 GB storage; Devboxes: 8 concurrent, 32 total); Team $40.00/month (1,000 MCUs ($50 at the $0.05 standard rate, i.e. $0.04 per MCU); MorphVM quota up to 256 vCPU, 1024 GB RAM, 4096 GB storage; Devboxes: 32 concurrent, 128 total); Scale $250.00/month (7,500 MCUs ($375 at the $0.05 standard rate, about $0.0333 per MCU); MorphVM quota up to 1024 vCPU, 4096 GB RAM, 16384 GB storage; Devboxes: 128 concurrent, 512 total) - Free tier: Conflicting: the pricing page lists the free Developer tier with 0 starting MCUs ('Purchase credits to use compute'), while the Plans doc says 300 MCUs free and the FAQ/Signing Up docs say 300 free compute hours per month. Access may require an invite. - Isolation: Virtual machine (MorphVM on Infinibranch; hypervisor not named) - Pause and resume: yes - Snapshots: yes - GPU: no - Custom images: yes - SDKs: Python, TypeScript, CLI, REST API - Regions: Not published - Open source: no - Self-hostable: no - Best for: Branching one running VM into many identical copies for parallel agent rollouts or RL environments; Pausing a VM with memory and processes intact and paying only snapshot storage while paused; Teams that can prepay a monthly MCU bundle for a 20-33% lower unit rate - Watch out for: Billing uses the largest of vCPU, RAM/4 GiB and disk/16 GiB, so a VM with a big disk or RAM pays for it even with 1 vCPU; Free tier is unclear: pricing page shows 0 MCUs while docs promise 300 free MCUs or compute hours; Egress and per-operation snapshot/instance charges are named in the terms but no rates are published; No published regions, network egress controls, or per-VM size limits - Sources: https://cloud.morph.so/web/pricing https://cloud.morph.so/docs/documentation/setup/plans https://cloud.morph.so/docs/documentation/setup/signing-up https://cloud.morph.so/docs/documentation/instances/pause-resume https://cloud.morph.so/web/legal/terms-of-service https://cloud.morph.so/docs/faq https://cloud.morph.so/docs/blog/developers https://www.morphllm.com/about - Facts checked: 2026-11-14 ## Freestyle Freestyle sells full Linux VMs aimed at long-running agent tasks and AI app builders, with commands, files, PTY sessions, domains, firewalls and VPCs managed through a TypeScript SDK and CLI. VMs can be paused with memory intact, snapshotted and started from snapshots, and resized up while running. vCPU, memory and storage are metered per second on the amount allocated, and every plan includes the same monthly compute allowance. Hobby and Pro have monthly minimums that count as usage credit. - Page: https://aisandboxhunt.com/freestyle - Company: Freestyle - Website: https://www.freestyle.sh - Categories: Dev boxes for AI agents, AI agent sandboxes - Score: 63/100, #18 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $50.00 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Price per vCPU-hour: $0.04032 - Price per GiB-hour of memory: $0.0129 - Storage per GiB-month: $0.06278 - Billing increment: per second (rates published per hour) - Billed while: wall-clock on allocated vCPU and memory while running, whether or not the guest is busy; storage is billed while the VM exists, including when paused - Plans: Free $0.00/month (Monthly allowance of 200 vCPU-hours, 400 GiB-hours memory, 60,000 GiB-hours storage, 50 GB transfer, as hard stops (VMs pause until reset); 10 concurrent VMs, 10 saved VMs, 10 snapshots; max 4 vCPU / 8 GiB / 32 GB per VM); Hobby $50.00/month (Same monthly compute allowance as Free plus $50 of usage covered by the fee; 500 GB transfer; 40 concurrent VMs, 200 saved VMs, 1,000 snapshots; max 8 vCPU / 16 GiB / 64 GB per VM); Pro $500.00/month (Same monthly compute allowance as Free plus $500 of usage covered by the fee; 500 GB transfer; 400 concurrent VMs, 4,000 saved VMs, 12,000 snapshots; max 32 vCPU / 64 GiB / 256 GB per VM) - Free tier: Free plan with a monthly allowance (same on every plan) of 200 vCPU-hours, 400 GiB-hours of memory, 60,000 GiB-hours of storage, plus 50 GB data transfer; no usage charges, allowances are hard stops. Adding a card charges a refunded $1 verification. - Isolation: Hardware-virtualized VM (supports nested virtualization; hypervisor not named) - Pause and resume: yes - Snapshots: yes - GPU: no - Custom images: yes - Network controls: yes - Max vCPU: 32 - Max memory GiB: 64 - SDKs: TypeScript, CLI - Regions: Not published - Open source: no - Self-hostable: no - Best for: Long-running agent workspaces that pause with memory intact and resume later; AI app builders that need per-user VMs with domains, firewalls and private VPC networking; Branching or cloning a VM state via cheap snapshots - Watch out for: Allocated vCPU and memory bill while running even when the guest is idle; set an idle timeout to pause; Host resources are oversubscribed; the vendor advises against sustained CPU-heavy work such as large builds; vm.exec() has a 5-minute maximum timeout; longer jobs need a PTY session; Resize is grow-only; to shrink a VM you must create a new one - Benchmark (ComputeSDK benchmarks, 2026-11-13): Burst start time (median of 100 at once) 1548 ms, rank 20 of 27 - Benchmark (aisandboxbench, 2026-10-23): Cold start (median) 0.693 s, rank 3 of 9 - Benchmark (aisandboxbench, 2026-10-23): Agent tool-call replay (median) 24.072 s, rank 6 of 9 - Sources: https://www.freestyle.sh/pricing https://www.freestyle.sh/docs/vms/pricing-and-limits https://www.freestyle.sh/docs/vms/lifecycle https://www.freestyle.sh/docs/vms https://www.freestyle.sh/docs/onboard.md - Facts checked: 2026-11-14 ## GitHub Codespaces GitHub Codespaces creates a development environment for a repository on a dedicated virtual machine, running a dev container that you open in VS Code, the browser or over the GitHub CLI. Compute is billed per hour while a codespace is active, in proportion to its cores: $0.18 an hour for 2 cores up to $2.88 for 32 cores, plus $0.07 per GB-month of storage while it exists. Personal GitHub Free accounts get 120 core-hours and 15 GB-months each month. Codespaces stop after an idle timeout and have a maximum lifetime of 12 hours before they must be restarted. - Page: https://aisandboxhunt.com/github-codespaces - Company: GitHub - Website: https://github.com/features/codespaces - Categories: Dev boxes for AI agents - Score: 63/100, #19 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $39.60 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Fixed-size price per hour: $0.18 (2-core machine) - Storage per GiB-month: $0.07 - Billing increment: Compute time reported hourly and billed monthly - Billed while: Compute while the codespace is active; storage for as long as the codespace or prebuild exists. - Plans: GitHub Free (personal account) $0.00/month (120 core-hours of compute and 15 GB-months of storage each month. Usage beyond it is billed if a payment method is on file, otherwise blocked) - Free tier: Personal accounts: 120 core-hours and 15 GB-months a month on GitHub Free, 180 core-hours and 20 GB-months on GitHub Pro. Organization plans include no free quota. - Isolation: Dedicated, private VM per codespace running a dev container - Pause and resume: yes - Custom images: yes - Max session hours: 12 - Max vCPU: 32 - SDKs: GitHub CLI, REST API, VS Code - Regions: Set automatically from your location, or chosen manually in settings; the region list is not on the pages checked. - Open source: no - Self-hostable: no - Best for: Repository-based dev environments defined with a dev container; Personal projects within the monthly free core-hours; Teams already billing through GitHub - Watch out for: Codespaces stop after at most 12 hours and need restarting; Organization accounts get no free quota; Per-core pricing is higher than most agent-focused dev box providers; Not designed as an API sandbox for agents - Sources: https://docs.github.com/en/billing/concepts/product-billing/github-codespaces https://docs.github.com/en/get-started/learning-about-github/githubs-plans https://docs.github.com/en/codespaces/about-codespaces/understanding-the-codespace-lifecycle https://docs.github.com/en/codespaces/about-codespaces/deep-dive - Facts checked: 2026-11-13 ## exe.dev exe.dev gives you persistent Linux VMs with root access, reachable with ssh exe.dev and published over HTTPS with built-in login, custom domains and secret-injecting integrations. The main product is a pool: you buy a fixed amount of vCPU and memory and run as many VMs in it as you like, from $15 a month for 2 vCPU / 4 GiB on Personal. Standalone sandbox VMs with their own reserved capacity are metered per second at $0.105 per 2 vCPU / 4 GiB hour on top of a plan. Disks sit on local NVMe and are copied to another server in the region for durability. - Page: https://aisandboxhunt.com/exe-dev - Company: exe.dev - Website: https://exe.dev - Categories: Dev boxes for AI agents - Score: 61/100, #21 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $38.10 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Fixed-size price per hour: $0.105 (Standalone VM unit: 2 vCPU, 4 GiB) - Storage per GiB-month: $0.08 - Billing increment: Per second, no minimum, no rounding up to the hour - Billed while: Standalone VMs: reserved vCPU and memory from creation until deletion, except while stopped (a stopped VM keeps its disk). Pools: billed for as long as the pool exists, empty or full. - Plans: Personal $15.00/month (One pool of 2 vCPU / 4 GiB for $15 a month (4 vCPU / 8 GiB $35, 8 vCPU / 16 GiB $75, 16 vCPU / 32 GiB $155). 50 VMs, 100 GiB of filesystem usage, 200 GiB of network bandwidth. Standalone VMs are metered on top); Work $75.00/month (Includes 2 vCPU + 4 GiB across all pools and standalone VMs; extra capacity $0.105 per 2 vCPU / 4 GiB hour. Unlimited pools in any region, 1,000 VMs per pool, 100 GiB disk per 2 vCPU, 50 GiB bandwidth per vCPU, unlimited users and SSO. Intro price $37.50 a month for the first 3 months) - Free tier: Personal plan free trial for 7 days. - Isolation: KVM-isolated VM per machine, each with its own VMM process - Pause and resume: yes - Custom images: yes - Network controls: yes - Max vCPU: 32 - Max memory GiB: 64 - SDKs: SSH CLI, HTTPS API, MCP server - Regions: Oregon, Los Angeles, New York, Dallas, Frankfurt, Tokyo, Sydney, Singapore and London (Oregon closed to new accounts). Personal pools use the account's region; Work pools can go in any region. - Open source: no - Self-hostable: no - Best for: Running many small long-lived VMs and apps on one flat monthly pool; SSH-first developers and agents who want root on a persistent Linux box; Disk-heavy builds that benefit from local NVMe - Watch out for: Pools are billed whether VMs run or not; only standalone VMs stop billing when stopped; Standalone sandbox VMs cost extra on top of the Personal subscription; Personal has one pool in the account's region and no team sharing; Memory is capped at 2 GiB per vCPU for standalone VMs - Sources: https://exe.dev/docs/standalone-vms.md https://exe.dev/pricing https://exe.dev/pricing/compare https://exe.dev/docs/pools.md https://exe.dev/docs/regions.md - Facts checked: 2026-11-13 ## Maritime Maritime offers two kinds of headless Linux micro-VM for agents. Sandboxes are short-lived machines billed per second at $0.0055 per vCPU-hour plus $0.00275 per GB-hour while running, ended with their disk after an idle period or a time limit of up to 24 hours. Agent boxes are persistent Firecracker micro-VMs that keep /data, sleep when idle and wake in about a second, for a flat $1 a month per 2 GB of memory with unlimited awake hours. Every plan fee becomes the same amount of usage credit, and usage beyond it is billed at the same rates. - Page: https://aisandboxhunt.com/maritime - Company: Maritime - Website: https://maritime.sh - Categories: Dev boxes for AI agents, AI agent sandboxes - Score: 61/100, #22 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $20.00 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Price per vCPU-hour: $0.0055 - Price per GiB-hour of memory: $0.00275 - Storage per GiB-month: $0.08 - Billing increment: per second, no minimum - Billed while: Sandboxes: wall-clock while running; a paused sandbox pays only for its disk. Agent boxes: a flat monthly price per 2 GB of memory, awake or asleep. - Plans: Starter $20.00/month ($20 of usage credit a month; 5 machines awake at once. Usage beyond the credit is billed at the same rates); Growth $100.00/month ($100 of usage credit a month; 25 machines awake at once; per-end-user cost reporting. Usage beyond the credit is billed at the same rates); Scale $500.00/month ($500 of usage credit a month; 60 machines awake at once; Swarms. Usage beyond the credit is billed at the same rates); Dedicated $5,000.00/month ($5,000 of usage credit a month; 500 machines awake at once; swarm runs up to 10,000 sandboxes; setup of apps, images and network by the vendor; BAA and DPA) - Free tier: 7-day free trial on Starter terms with up to $20 of usage. A card is needed to start the first machine but is not charged unless the trial is not cancelled, in which case $20 is charged on day 7 as that month's credit. - Isolation: Micro-VM per machine with its own kernel and disk (Firecracker for agent boxes) - Pause and resume: yes - Snapshots: yes - Custom images: yes - Network controls: yes - Max vCPU: 16 - Max memory GiB: 64 - SDKs: TypeScript, Python, CLI, REST API, MCP, E2B-compatible API - Regions: Not stated on the pages checked. - Open source: no - Self-hostable: no - Best for: One persistent machine per end user or per agent at a flat low monthly price, with unlimited awake hours; Short-lived code execution at low per-second rates, including existing E2B SDK code pointed at Maritime; Mixing sandboxes and persistent boxes under one plan credit and invoice - Watch out for: Agent boxes are billed every month whether awake or asleep, and self-serve sizes stop at 8 GB of memory; Sandboxes end with their disk after a few idle minutes by default; background commands do not count as activity; The plan limits how many machines can be awake at once (5 on Starter); further starts are refused; Usage beyond the plan credit has no spending cap - Sources: https://maritime.sh/pricing https://maritime.sh/docs/ephemeral-sandboxes.md https://maritime.sh/docs/persistent-sandboxes.md https://maritime.sh - Facts checked: 2026-11-14 ## Namespace Devboxes Namespace Devboxes are remote Linux, macOS and Windows machines on Namespace's own hardware, reachable over SSH, IDEs, a CLI, a TypeScript or Go SDK and MCP. Linux Devboxes run from a container image you choose and include Docker; macOS Devboxes run on Apple Silicon. Persistent Devboxes keep their files across stops and start again when you connect; ephemeral ones are also available. Compute is billed in Devbox Minutes at $0.004 each (a small Linux Devbox uses one per minute) plus $0.20 per GB-month of storage actually used. - Page: https://aisandboxhunt.com/namespace - Company: Namespace Labs - Website: https://namespacelabs.com/devbox - Categories: Dev boxes for AI agents, AI agent sandboxes, CI runners - Score: 61/100, #23 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $52.80 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Fixed-size price per hour: $0.24 (Linux size S: up to 4 vCPU, 8 GB RAM (1 Devbox Minute per minute at $0.004). M, L and XL use 2, 4 and 8 Devbox Minutes per minute.) - Storage per GiB-month: $0.20 - Billing increment: per minute, in Devbox Minutes - Billed while: Compute for each minute the Devbox runs; persistent storage for data actually stored, including while stopped. - Plans: Developer $0.00/month (Pay as you go. 1 Devbox per user, 3 per workspace (running and stopped both count). Max Devbox duration 3 hours); Devbox Plus $250.00/month ($250/month including 50,000 Devbox Minutes and 400 GB-month of storage. 5 Devboxes per user, 50 per workspace. Max Devbox duration 3 hours); Team $100.00/month ($100/month; its included unit minutes cover CI and remote builds, not Devboxes, which are billed separately. 3 Devboxes per user, 30 per workspace. Max Devbox duration 5 hours); Business $250.00/month ($250/month; its included unit minutes cover CI and remote builds, not Devboxes, which are billed separately. 5 Devboxes per user, 50 per workspace. Max Devbox duration 24 hours) - Free tier: 30-day free trial of the Developer plan on sign-up, including Devboxes; no credit card required. The pricing page does not state how much usage the trial covers. - Isolation: Namespace compute instance per Devbox on Namespace-operated hardware; the pages checked do not name the virtualization technology - Pause and resume: yes - GPU: no - Custom images: yes - Network controls: yes - Max session hours: 24 - Max vCPU: 32 - Max memory GiB: 64 - SDKs: TypeScript, Go, CLI, MCP server - Regions: Runs in Namespace's own sites; the pages checked do not list them (the CLI can measure latency to each Devbox site). - Open source: no - Self-hostable: no - Best for: Coding agents that need a full dev machine with Docker, the repository checked out and a shared environment definition; Agents that must build or test on macOS (Apple Silicon with Xcode) or Windows as well as Linux; Teams already using Namespace for CI who want dev machines in the same account - Watch out for: Each run is capped at 3 hours on Developer and Devbox Plus (5 on Team, 24 on Business); Low Devbox count limits: 1 per user and 3 per workspace on the pay-as-you-go plan, counting stopped ones; macOS Devboxes cost 15 times the Linux rate per size step; Team and Business plan fees do not include any Devbox usage - Benchmark (ComputeSDK benchmarks, 2026-11-13): Repo clone, install and typecheck (one run) 47445 ms, rank 3 of 25 - Benchmark (StarSling HPC sandbox benchmarks, 2026-09-29): Node.js web tooling (runs per second) 21.79 runs/s, rank 2 of 12 - Benchmark (StarSling HPC sandbox benchmarks, 2026-09-29): Better Auth cold dependency install 11.4 s, rank 1 of 11 - Sources: https://namespacelabs.com/pricing https://namespace.so/docs/devbox/lifecycle https://namespace.so/docs/devbox https://namespace.so/docs/security/egress-policy https://namespacelabs.com - Facts checked: 2026-11-14 ## Aether Aether runs each coding-agent task in its own Firecracker microVM, an Ubuntu workspace with persistent storage, a terminal, a browser and SSH access. You bring Claude Code, Codex or OpenCode on your own subscription or API key (or buy model calls separately), and Aether bills the VM in credits for active runtime only, per second. Workspaces suspend when idle and resume where they left off. Plans grant monthly credits, and credit packs cost $0.10 per credit. - Page: https://aisandboxhunt.com/aether - Company: Aether - Website: https://www.runaether.dev - Categories: Dev boxes for AI agents, AI agent sandboxes - Score: 60/100, #25 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $95.00 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Fixed-size price per hour: $1.00 (Medium: 2 vCPU / 8 GB / 20 GB volume (10 credits per hour at $0.10 per credit)) - Billing increment: per second (credits prorated by the second) - Billed while: Only while the workspace VM is active. Idle and suspended workspaces are not billed; a workspace suspends once no task, background activity or interactive connection holds it. - Plans: Free $0.00/month (20 credits a month (about 2 hours of Medium, or about 4 hours of the Small size listed on the homepage), 1 warm slot, 3 tasks at once); Starter $30.00/month (400 credits a month (about 40 hours of Medium), 2 warm slots, 10 tasks at once); Pro $75.00/month (2,000 credits a month (about 200 hours of Medium), 3 warm slots, 25 tasks at once); Max $120.00/month (4,000 credits a month (about 400 hours of Medium), 5 warm slots, 50 tasks at once) - Free tier: Free plan: 20 credits a month at no cost, permanently (about 2 hours of the 2 vCPU / 8 GB Medium size). - Isolation: Firecracker microVM per task, not reused across tasks or users - Pause and resume: yes - Snapshots: yes - GPU: no - Max vCPU: 8 - Max memory GiB: 32 - SDKs: CLI, REST API, WebSocket API, MCP - Regions: Not stated on the pages checked. - Open source: no - Self-hostable: no - Best for: Running Claude Code, Codex or OpenCode on your own subscription in a cloud VM you can watch, steer and SSH into; Task-per-VM agent work tied to GitHub pull requests, Slack, Linear or schedules; Bursty agent work, since idle and suspended workspaces cost nothing - Watch out for: At the $0.10 credit-pack price a 2 vCPU / 8 GB VM costs $1.00 per hour; plan credits are cheaper but expire monthly; Workspaces are tied to a task and, per the homepage, are deleted when its pull request merges or closes, or after 48 hours of inactivity; Only three fixed sizes in the docs, up to 8 vCPU / 32 GB; no GPU; The homepage and docs disagree on the default size and whether a 1 vCPU Small size exists - Sources: https://www.runaether.dev https://docs.runaether.dev/concepts/pricing-and-tiers https://docs.runaether.dev/concepts/billing-and-credits https://docs.runaether.dev/concepts/workspaces - Facts checked: 2026-11-14 ## Orgo Orgo provides full desktop computers for AI agents: each Linux or Windows computer is its own virtual machine with a disk, desktop, browser and terminal, controlled through screenshots, clicks, typing and bash over an API, SDKs or MCP. Computers run until you stop them, with no time limit, and are backed up daily. Linux computers can run by the hour from credit, from $0.0548 an hour for a Medium (2 vCPU, 8 GB), or on a monthly plan that includes shared vCPU and RAM. Windows runs only on a plan; Mac and iPhone hardware is separate. - Page: https://aisandboxhunt.com/orgo - Company: Orgo - Website: https://www.orgo.ai - Categories: Dev boxes for AI agents, Browser sandboxes for AI agents - Score: 60/100, #26 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $36.05 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Fixed-size price per hour: $0.05479 (Medium Linux: 2 vCPU, 8 GB RAM, 40 GB disk) - Price per browser-hour: $0.02312 - Storage per GiB-month: $0.05 - Billing increment: Per second, no hourly minimum - Billed while: By the hour: the running price covers vCPU, RAM and disk while running; a stopped computer pays only for its disk. Plan capacity is a flat monthly charge. - Plans: Starter $29.00/month (2 vCPU and 8 GB of shared capacity for computers on the plan, $5 of monthly AI & Platform credit, 1 GB residential IP data, 5 people per workspace, up to 50 computers); Growth $99.00/month (8 vCPU and 32 GB of shared capacity, $10 of monthly AI & Platform credit, 5 GB residential IP data, 5 people per workspace, up to 50 computers); Scale $399.00/month (36 vCPU and 128 GB of shared capacity, $50 of monthly AI & Platform credit, 20 GB residential IP data, 25 people per workspace, Windows included, up to 50 computers) - Free tier: Free account: $5 starting credit (or $2 for an agent signup) that runs one Medium Linux computer by the hour. - Isolation: Isolated virtual machine per computer - Pause and resume: yes - Max vCPU: 6 - Max memory GiB: 32 - SDKs: Python, TypeScript, CLI, MCP server, REST API - Regions: Not stated - Open source: no - Self-hostable: no - Best for: Computer-use agents that need a full Linux or Windows desktop rather than a browser tab; Long-running agent desktops with no session time limit; Watching or taking over an agent's screen live - Watch out for: Hourly use is 25% dearer than plan capacity for always-on computers; Windows needs a plan and costs extra below Scale; Free accounts run one Medium computer at a time and other sizes need a plan; Regions and network controls are not documented on the pages checked - Sources: https://docs.orgo.ai/guides/pricing https://docs.orgo.ai/guides/billing https://docs.orgo.ai/guides/instance-types https://docs.orgo.ai/llms.txt https://www.orgo.ai/pricing - Facts checked: 2026-11-15 ## Tembo Tembo runs coding agents such as Claude Code, Codex, Cursor, Amp and OpenCode on tasks from GitHub, Linear, Slack, Sentry and other tools, with each session in a dedicated Linux VM. VM compute is metered per second at $0.0403 per vCPU-hour plus $0.0130 per GiB-hour while the VM runs, and draws from a monthly dollar allowance that also covers model calls through Tembo's own gateway. Paid plans include more usage than their fee and allow optional overage billing with a spend limit. Sandboxes are set up per session and torn down when it ends, apart from cached project environments. - Page: https://aisandboxhunt.com/tembo - Company: Tembo - Website: https://www.tembo.io - Categories: Dev boxes for AI agents, Coding agents with cloud machines - Score: 58/100, #27 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $60.00 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Price per vCPU-hour: $0.0403 - Price per GiB-hour of memory: $0.013 - Billing increment: per second - Billed while: Only while the session VM is running. Stopped, snapshotted and otherwise non-running VMs accrue no compute. - Plans: Pro $60.00/month ($90 of usage a month shared between VM compute and Tembo Gateway model usage; up to 5 users; VM size up to Medium (4 vCPU / 8 GiB); optional overage billing with a spend limit); Max $200.00/month ($400 of usage a month shared between VM compute and Tembo Gateway model usage; up to 10 users; VM size up to XL (16 vCPU / 32 GiB); optional overage billing with a spend limit) - Free tier: 7-day free trial of the Max plan with an included usage allowance (amount not stated), up to 10 users and VM size up to XL. Sessions stop when the trial or its allowance ends; no overage billing during the trial. - Isolation: Dedicated Linux VM per session, with nested virtualization for Docker - Snapshots: yes - Max vCPU: 16 - Max memory GiB: 32 - SDKs: TypeScript, REST API, MCP - Regions: Not stated on the pages checked; self-hosted and VPC deployment are offered for organizations. - Open source: no - Self-hostable: yes - Best for: Teams that want agents triggered from issue trackers, alerts and chat, with each session in its own VM; Comparing or switching between several coding agents while paying one per-second VM rate; Workloads that need Docker inside the sandbox, since VMs support nested virtualization - Watch out for: VMs are only available inside Tembo sessions; there is no standalone sandbox product; Pro limits VMs to 4 vCPU / 8 GiB; larger sizes need Max; The monthly allowance is shared with model usage if you use Tembo's gateway instead of your own key or subscription; Sandboxes are torn down at the end of a session, apart from cached project environments - Sources: https://www.tembo.io/pricing https://docs.tembo.io/resources/pricing https://docs.tembo.io/features/sandbox/overview https://www.tembo.io - Facts checked: 2026-11-14 ## Islo Islo, from the team behind Incredibuild, runs coding agents on persistent Linux machines, each in its own microVM, with the repository, dependencies, browser and tools prepared in advance. Agents can be triggered from Slack, GitHub or Linear, and a gateway injects credentials and enforces allow or deny rules on outbound traffic. Sandboxes pause on idle or on a runtime cap, keep their files, and can resume on the next request. Plans cost $20 a month for one developer or $200 for a team and include sandbox compute that refreshes weekly; extra time is paid from credits at $0.07 per CPU-hour and $0.04 per GB-hour of provisioned memory. - Page: https://aisandboxhunt.com/islo - Company: Islo - Website: https://islo.dev - Categories: Dev boxes for AI agents, Coding agents with cloud machines - Score: 57/100, #28 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $86.00 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Price per vCPU-hour: $0.07 - Price per GiB-hour of memory: $0.04 - Storage per GiB-month: $0.511 - Billing increment: Not stated; rates are quoted per hour - Billed while: Only while the sandbox runs, from plan compute or credits: CPU per vCPU, memory per GB provisioned, and disk. No idle fees. - Plans: Individual $20.00/month (Sandbox compute included (amount not stated), refreshed weekly. Up to 5 sandboxes at once, 1 member. Extra time is paid from credits); Team $200.00/month (10x the compute of Individual. Up to 50 sandboxes at once, sandboxes up to 32 vCPU and 32 GB memory, unlimited members) - Free tier: 14-day free trial with a work email and no card; nothing is charged automatically after it. - Isolation: microVM per Islo Machine - Pause and resume: yes - Snapshots: yes - Custom images: yes - Network controls: yes - Max vCPU: 32 - Max memory GiB: 32 - SDKs: Python, TypeScript, CLI, REST API - Regions: Not stated; the API lists the compute regions allowed for an account. Bring-your-own-cloud deployment is offered. - Open source: no - Self-hostable: no - Best for: Teams handing tickets from Slack, GitHub or Linear to coding agents that work on a prepared machine; Running agents without giving them raw API keys, since the gateway injects credentials; Agent evaluations that need isolated machines with egress control - Watch out for: CPU and memory are billed on provisioned size while running, at higher rates than many sandbox providers; Regions and the amount of compute each plan includes are not published; The product is organized around its own factories, agents and lines, which adds concepts beyond a plain VM - Sources: https://islo.dev/pricing https://islo.dev/security/ https://islo.dev/ https://docs.islo.dev/concepts/sandbox-lifecycle.md https://docs.islo.dev/llms.txt - Facts checked: 2026-11-15 ## Bitrise Remote Dev Environments Bitrise Remote Dev Environments are on-demand macOS (Apple silicon) and Linux VMs running on the same stacks and caches as Bitrise CI, with iOS simulators and Android emulators available in a session. Sessions keep a persistent disk when terminated and can be restored later; they auto-terminate after five days of inactivity by default, configurable up to 30 days or off. Plans include minutes per operating system, then bill per minute up to a monthly ceiling, after which usage is free for the rest of the month. Agents connect through an MCP server, the CLI, an HTTP API or SSH. - Page: https://aisandboxhunt.com/bitrise-rde - Company: Bitrise - Website: https://bitrise.io/platform/remote-dev-environments - Categories: Dev boxes for AI agents - Score: 53/100, #30 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $75.97 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Fixed-size price per hour: $0.264 (Linux VM up to 4 vCPU / 16 GB, Basic plan overage rate ($0.0044 per minute)) - Billing increment: per minute - Billed while: While a session is running, after the plan's included minutes for that operating system; billing stops for the month once that operating system's monthly ceiling is reached. Terminated sessions keep their disk at no separate charge. - Plans: Basic Tier 1 $20.00/month (480 included minutes a month per OS (8 hours macOS and 8 hours Linux), 1 concurrent VM. macOS M2 Pro Large (6 vCPU / 14 GB), Linux up to 4 vCPU / 16 GB. Overage macOS $0.0288/min, Linux $0.0044/min; monthly ceiling macOS $99, Linux $39. $18/month billed annually); Basic Tier 2 $40.00/month (960 included minutes a month per OS, 3 concurrent VMs. Same machines and overage rates as Basic Tier 1; monthly ceiling macOS $297, Linux $117. $36/month billed annually); Basic Tier 3 $60.00/month (1,440 included minutes a month per OS, 5 concurrent VMs. Same machines and overage rates as Basic Tier 1; monthly ceiling macOS $495, Linux $195. $54/month billed annually); Pro Tier 1 $40.00/month (480 included minutes a month per OS, 1 concurrent VM. macOS M2 Pro X Large (12 vCPU / 28 GB), Linux up to 8 vCPU / 32 GB. Overage macOS $0.0576/min, Linux $0.0088/min; monthly ceiling macOS $199, Linux $79. $36/month billed annually); Pro Tier 2 $80.00/month (960 included minutes a month per OS, 3 concurrent VMs. Same machines and overage rates as Pro Tier 1; monthly ceiling macOS $597, Linux $237. $72/month billed annually); Pro Tier 3 $120.00/month (1,440 included minutes a month per OS, 5 concurrent VMs. Same machines and overage rates as Pro Tier 1; monthly ceiling macOS $995, Linux $395. $108/month billed annually) - Free tier: 14-day free trial with Tier 1 allocation (8 hours per OS, 1 concurrent VM). A credit card is required at signup for verification but is not charged, the trial does not convert to a paid plan automatically, and trial sessions are limited to eight hours. - Isolation: Full macOS or Linux VM per session - Pause and resume: yes - Custom images: no - Network controls: no - Max vCPU: 12 - Max memory GiB: 32 - SDKs: CLI, MCP server, HTTP API (experimental), VS Code extension - Regions: Not stated on the pages checked; enterprise customers can get dedicated or private cloud environments. - Open source: no - Self-hostable: no - Best for: Mobile teams whose agents need Xcode on real macOS, iOS simulators or Android emulators; Persistent per-task dev machines that can be terminated and restored later with their disk; Teams already on Bitrise CI that want agents to work on the same stacks and caches as their builds - Watch out for: Plans include only 8 to 24 hours a month per operating system; usage beyond that is billed per minute until the monthly ceiling; Machine sizes are fixed per plan: Basic tops out at a 4 vCPU Linux VM and a 6 vCPU Mac; Egress filtering and custom VM images are not available yet; Session disks are not backed up and are lost if the underlying stack is removed - Sources: https://bitrise.io/pricing https://bitrise.io/platform/remote-dev-environments https://docs.bitrise.io/en/bitrise-rde/getting-started/rde-faq - Facts checked: 2026-11-14 ## Coder Coder is an AGPL-3.0 platform you deploy yourself to provision development workspaces for people and coding agents. Workspaces are defined in Terraform templates, so they can be VMs, Kubernetes pods or Docker containers, and idle ones shut down automatically. The free Community edition has unlimited workspaces and templates; Premium and AI Premium add enterprise access control, high availability, support and unlimited Coder Agents, priced per user on request. - Page: https://aisandboxhunt.com/coder - Company: Coder Technologies - Website: https://coder.com - Categories: Dev boxes for AI agents - Score: 50/100, #31 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: Self-host · no list price for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Billing increment: Not applicable: free software; you pay your own infrastructure costs - Billed while: Not applicable: workspaces cost whatever your cloud or servers charge - Free tier: Community edition is free and open source (AGPL-3.0) with unlimited workspaces and templates, up to 5 active Coder Agents and 1 platform integration. - Isolation: Depends on the template: cloud VMs, Kubernetes pods or Docker containers provisioned with Terraform - Pause and resume: yes - Custom images: yes - Network controls: yes - SDKs: CLI, REST API, Terraform provider, VS Code and JetBrains plugins - Regions: Wherever you deploy it: your own cloud accounts, Kubernetes clusters or servers. - Open source: yes (AGPL-3.0) - Self-hostable: yes - Best for: Companies that must keep developer and agent environments inside their own cloud or data center; Standardizing workspaces as Terraform templates across VMs, Kubernetes and Docker; Running Claude Code, Codex and other agents in governed, auto-stopping workspaces - Watch out for: Paid tiers have no public price; You operate the control plane, database and infrastructure; Isolation strength depends on the template you choose (container or VM) - Sources: https://github.com/coder/coder https://coder.com/pricing https://github.com/coder/coder/releases/tag/v2.38.7 - Facts checked: 2026-11-13 ## Containarium Containarium is an Apache-2.0 agent runtime that gives each agent a persistent, SSH-reachable Linux box in an unprivileged container, managed through a CLI and MCP servers. It runs on LXC or Kubernetes, offers an opt-in per-tenant eBPF egress policy, a model gateway that keeps provider keys out of the box, and GPU passthrough. Self-hosting is free on your own Ubuntu VM; a hosted cloud sells a $20 per month seat with a fixed CPU, memory, disk and transfer allowance and no overage billing. - Page: https://aisandboxhunt.com/containarium - Company: FootprintAI - Website: https://containarium.dev - Categories: Dev boxes for AI agents - Score: 48/100, #32 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: Self-host · no list price for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Billing increment: Self-hosted: not applicable. Hosted: flat monthly seat, not metered - Billed while: Self-hosted: no charges. Hosted: flat per-seat monthly price regardless of usage - Plans: Seat (hosted) $20.00/month (One box per seat: 2 vCPU (1 guaranteed, bursts to 2 when the host has capacity), 4 GiB RAM, 20 GB disk, 100 GB transfer per month. No metering and no overage charges; idle boxes sleep. Marked as launch pricing) - Free tier: Self-hosting is free under Apache-2.0. The hosted cloud has no free tier listed. - Isolation: Unprivileged LXC container (or Kubernetes pod) per box; tenant-to-tenant isolation is not on by default - GPU: yes - Network controls: yes - SDKs: CLI, MCP server, SSH - Regions: Self-hosted: your own Ubuntu 24.04 VM or Kubernetes cluster. Hosted cloud region not stated. - Open source: yes (Apache-2.0) - Self-hostable: yes - Best for: Giving coding agents a persistent Linux box they reach over SSH and MCP; Self-hosting many agent boxes on one Ubuntu VM or a Kubernetes cluster; Users who want a flat monthly price with no usage bill on the hosted cloud - Watch out for: Container isolation, and tenant-to-tenant isolation is off by default; Hosted seat is a shared box with only 1 guaranteed vCPU; Hosted price is launch pricing and may change - Sources: https://github.com/FootprintAI/Containarium https://containarium.dev/pricing https://github.com/FootprintAI/Containarium/releases/tag/v0.136.0 - Facts checked: 2026-11-13 ## Okteto Okteto runs development, preview, testing and coding-agent environments as containers on a Kubernetes cluster you own, syncing local code into a remote development container. The Starter plan is free for up to 5 users with 500 environments a month; Scale and Enterprise add SSO, access control, cost controls, SLAs and BYOC, priced per user on request. The Okteto CLI is Apache-2.0, but the platform itself is commercial software installed with a license. - Page: https://aisandboxhunt.com/okteto - Company: Okteto - Website: https://www.okteto.com - Categories: Dev boxes for AI agents - Score: 47/100, #33 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: Self-host · no list price for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Billing increment: Not applicable: priced per user, not per environment or minute - Billed while: Not applicable: you pay your own cluster costs - Plans: Starter $0.00/month (Free for up to 5 users and 500 environments a month, one-year license, self-hosted, community forum support. No SSO, RBAC or SLA) - Free tier: Starter: free self-hosted license for up to 5 users and 500 environments a month, valid for one year, no credit card needed. - Isolation: Development containers in Kubernetes namespaces on your own cluster - Custom images: yes - SDKs: CLI, Okteto manifest - Regions: Your own Kubernetes cluster (EKS, GKE, AKS, Nutanix NKP or bare metal); BYOC on AWS or GCP on Enterprise. - Open source: no - Self-hostable: yes - Best for: Small teams (up to 5 users) wanting free self-hosted dev and preview environments on Kubernetes; Running coding agents in isolated environments on a cluster you already own; Developing against realistic copies of Kubernetes services - Watch out for: Starter is limited to 5 users and 500 environments a month and lacks SSO and RBAC; Paid tiers have no public price; Requires and runs on your own Kubernetes cluster; environments are containers - Sources: https://www.okteto.com/pricing/ https://github.com/okteto/okteto https://github.com/okteto/okteto/releases/tag/3.25.0 - Facts checked: 2026-11-15 ## Box by Simular Box by Simular gives an AI agent a cloud Linux computer of its own: an Ubuntu desktop with Chrome, files and a persistent shell, driven through one hosted MCP server by Claude, ChatGPT, Gemini, Claude Code, Codex, Cursor or any MCP client. Each account gets a box that sleeps when idle, wakes in about a second with files and logins kept, and can be watched live and taken over in the browser. API keys can be stored so commands on the box use them while the agent only sees a stand-in. The box itself is free; the computer-use model calls it makes draw on the Simular account's credits. A box left asleep for 14 days in a row is deleted. - Page: https://aisandboxhunt.com/simular-box - Company: Simular - Website: https://box.simular.ai - Categories: Dev boxes for AI agents, Browser sandboxes for AI agents - Score: 46/100, #34 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: No public rate for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Billing increment: Not stated; the box itself is free - Billed while: The box has no charge of its own; the model calls it makes use Simular account credits. - Plans: Free $0.00/month (One cloud Linux desktop per account at no charge; model calls use the account's credits, which can be topped up or raised with a Simular plan) - Free tier: The box itself is free. Simular's free account plan refreshes credits daily for model calls; the daily amount is not stated. - Isolation: Not stated; a cloud Ubuntu desktop per account - Pause and resume: yes - SDKs: MCP server, CLI - Regions: Not stated - Open source: no - Self-hostable: no - Best for: Giving Claude, ChatGPT or a coding agent a real desktop and browser through a single remote MCP server; One shared computer that several agents use, with logins and files kept between tasks; Watching an agent's screen live and stepping in for sign-ins, 2FA codes or payments - Watch out for: Use is through the MCP server only; there is no SDK or REST API for creating machines from code; Machine size, isolation and regions are not documented; Model credits per task are not published, and pricing may change after the beta; An unused box is deleted after 14 days asleep - Sources: https://box.simular.ai/llms.txt https://box.simular.ai/docs/mcp.md https://box.simular.ai/docs https://box.simular.ai/docs/secrets.md https://www.simular.ai/pricing https://www.simular.ai/articles/linux-desktop-in-the-cloud - Facts checked: 2026-11-15 ## boxes.dev boxes.dev gives each coding agent its own cloud Linux machine (a devbox) created from a Template box snapshot of your dev environment, including memory. You bring your own Codex or Claude Code subscription or API key; boxes.dev bills only machine time, measured in box-hours, where one box-hour is one hour awake on the standard 4 vCPU / 8 GiB box. Devboxes sleep after about five minutes without activity, keeping files and memory, and sleeping boxes are not billed. Paid seats include a monthly box-hour allowance, with extra usage at $0.40 per box-hour from prepaid credit. - Page: https://aisandboxhunt.com/boxes-dev - Company: boxes.dev - Website: https://boxes.dev - Categories: Dev boxes for AI agents - Score: 45/100, #35 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: $87.00 for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Fixed-size price per hour: $0.40 (Standard box: 4 vCPU / 8 GiB RAM / 100 GB disk (one box-hour per awake hour)) - Billing increment: per second - Billed while: Only while the box is awake. Sleeping devboxes and Template boxes use no box-hours. A box sleeps after about 5 minutes without activity by default (5, 15, 30 or 60 minutes selectable); a running agent or a keep-awake request keeps it awake. - Plans: Starter $19.00/month (Per paid seat: 50 box-hours a month, 5 awake devboxes plus the Template box, up to 5 projects, email support. Extra usage $0.40 per box-hour from prepaid credit); Pro $99.00/month (Per paid seat: 300 box-hours a month, 12 awake devboxes plus the Template box, up to 10 projects, priority support. Extra usage $0.40 per box-hour from prepaid credit); Max $249.00/month (Per paid seat: 750 box-hours a month, 25 awake devboxes plus the Template box, up to 20 projects, priority support. Extra usage $0.40 per box-hour from prepaid credit) - Free tier: Trial: free, 20 box-hours per user (worth $8 at the extra-usage rate), 3 awake devboxes plus the Template box, up to 2 projects. It is a one-time trial per user that starts at first machine use; some teams must add a payment method first. - Isolation: Separate Linux machine per devbox, with its own filesystem, processes and ports (virtualization technology not stated) - Pause and resume: yes - Snapshots: yes - Custom images: yes - Max vCPU: 8 - Max memory GiB: 32 - SDKs: CLI, REST API, Desktop app, iOS and Android apps - Regions: Not stated on the pages checked. - Open source: no - Self-hostable: no - Best for: Running several Claude Code or Codex sessions in parallel, each on its own copy of a full dev environment; Developers who already pay for a Claude or ChatGPT plan and only want to pay for machine time; Long-lived devboxes that sit idle much of the time, since sleeping boxes keep memory and files and are not billed - Watch out for: Compute is sold in fixed box sizes up to 8 vCPU / 32 GiB, not per vCPU or per GiB; Seat fees are per user, and each seat's included box-hours cannot be shared beyond the team's prepaid credit pool; The product is built around its agent workbench (desktop, mobile, Slack, Linear); there is no general-purpose sandbox SDK; Awake time counts even when idle if a keep-awake request or a longer sleep timeout is set - Sources: https://boxes.dev https://boxes.dev/help/plans-and-box-hours https://boxes.dev/help/sleep-wake-recover https://boxes.dev/help/data-and-credentials - Facts checked: 2026-11-14 ## OpenComputer OpenComputer runs coding agents such as Claude Code and Codex in its cloud: you mention @Computer in Slack, or deploy your own TypeScript agent and call it over HTTP. Each agent works in a sandbox with your repository, dependencies and build caches kept warm between tasks. The machine is not sold separately: plans cover a Slack workspace and include work minutes, counted only while the agent is working, and you bring your own model subscription or pay for tokens at API prices. The code is open source under Apache-2.0 and can be self-hosted for free on one VM per organization in AWS, Azure or Google Cloud. - Page: https://aisandboxhunt.com/opencomputer - Company: OpenComputer - Website: https://opencomputer.dev - Categories: Dev boxes for AI agents, Coding agents with cloud machines - Score: 41/100, #37 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: Self-host · no list price for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Billing increment: Work minutes metered by the second - Billed while: Only while the agent is working, from picking up a message to its reply; idle threads and time waiting on people are free. - Plans: Pro $20.00/month (2,000 work minutes a month for the whole Slack workspace. Extra minutes $0.01 each from prepaid credits. Model use on your own subscription or key, or on OpenComputer's tokens at API prices); Max $200.00/month (25,000 work minutes a month for the whole Slack workspace. Extra minutes $0.01 each from prepaid credits) - Free tier: 3-day free trial with 600 work minutes and $10 of model tokens, no card; it ends at 600 minutes or day 3, whichever comes first. - Isolation: Not stated - Network controls: yes - SDKs: TypeScript, CLI, REST API, Slack - Regions: Usually the United States by default; an agent can be pinned to US, EU, Asia-Pacific and narrower regions including Canada, UK, Japan, Australia, Middle East, South America and Africa. - Open source: yes - Self-hostable: yes - Best for: Teams that want to hand small code changes to an agent from a Slack thread; Using an existing Claude or ChatGPT subscription for cloud agent work; Deploying custom agents as HTTP endpoints, on schedules or webhooks; Teams that want to run the same Slack agent free on their own cloud account - Watch out for: No general-purpose machine or SSH dev box is sold; access is through the agent; Sandbox size, isolation and limits are not published; Plans are per Slack workspace, and parallel threads use minutes in parallel - Sources: https://opencomputer.dev https://docs.opencomputer.dev/self-hosting/overview https://docs.opencomputer.dev/agents/regions.md https://docs.opencomputer.dev/agents/egress.md - Facts checked: 2026-11-14 ## Replit Replit is a browser-based development platform built around Replit Agent, which builds apps from prompts in a hosted cloud workspace and publishes them with databases and storage. The development machine is not priced on its own: plans include agent usage and monthly credits, and publishing, databases and storage are billed from those credits. Core costs $20 a month with $20 of credit for more powerful models; Pro starts at $100 a month with $100 of credits and up to 15 collaborators. The free Starter plan gives daily agent credits and one published app. - Page: https://aisandboxhunt.com/replit - Company: Replit - Website: https://replit.com - Categories: Dev boxes for AI agents, Coding agents with cloud machines - Score: 38/100, #38 of 38 dev boxes for AI agents (method 1.3 (2026-11-14)) - List price per month: No public rate for 2 vCPU, 4 GiB, 220 hours a month (10 h × 22 days), CPU busy 20%, paused the rest of the time - Billing increment: Not stated - Billed while: The development workspace is included in the plan; publishing, databases, storage and paid agent modes draw on credits. - Plans: Starter $0.00/month (Daily agent credits up to a monthly cap, monthly credits for databases, storage and publishing, Lite build only, 1 published app that goes down after 30 days); Core $20.00/month ($20 a month ($18 billed yearly). Up to 30 hours of chat and 60 projects on Free Mode, $20 toward more powerful models, unlimited published apps, up to 5 collaborators, 1 active background task. Pay-as-you-go beyond credits); Pro $100.00/month ($100 a month ($90 billed yearly) with $100 of credits; higher credit tiers up to $2,000 a month. Up to 15 collaborators, 10 parallel tasks, credit rollover for two months, 28-day database rollback, premium support) - Free tier: Starter plan: free daily agent credits up to a monthly cap and monthly credits for cloud usage, with one published app. - Isolation: Not stated - SDKs: Web IDE, Mobile app - Regions: Not stated - Open source: no - Self-hostable: no - Best for: Building and publishing apps from prompts without setting up a machine; Non-developers and small teams prototyping in the browser; Projects that want hosting, databases and the editor in one place - Watch out for: No raw machine to rent; workspace resources are not published; Spending is driven by agent credits, which can run out quickly on heavy use; Not an API sandbox for running your own agents; Free published apps go down after 30 days - Sources: https://replit.com/pricing https://docs.replit.com/billing/plans/replit-core https://docs.replit.com/billing/plans/starter-plan https://docs.replit.com/billing/plans/replit-pro - Facts checked: 2026-11-14 # Browser sandboxes for AI agents Hosted browsers and desktops that agents control to browse, click and fill forms. Priced per browser-hour, so they are compared only with each other. Ranking: https://aisandboxhunt.com/browser-sandboxes ## Kernel Kernel runs each Chromium browser in its own unikernel microVM, driven with Playwright, CDP or computer controls. Usage is billed per second at $0.0000166667 per GB-second: a 1 GB headless browser costs about $0.06 per hour and the default 8 GB headful browser about $0.48 per hour. Browsers go into standby about 10 seconds after the last client disconnects, which stops usage charges, and proxies are not charged. Every plan includes monthly usage credit, starting with $5 on the free plan. - Page: https://aisandboxhunt.com/kernel - Company: Kernel - Website: https://www.kernel.sh - Categories: Browser sandboxes for AI agents - Score: 76/100, #3 of 12 browser sandboxes for AI agents (method 1.3 (2026-11-14)) - Per browser-hour: $0.06 - Price per GiB-hour of memory: $0.06 - Price per browser-hour: $0.06 - GPU prices per hour: Headful browser with GPU acceleration $2.88 - Billing increment: Per second - Billed while: Only while the browser is active; standby (about 10 seconds after the last client disconnects) stops charges, and idle browsers in a pool are not charged - Plans: Developer $0.00/month ($5 of usage credit per month; 5 concurrent browsers; stealth mode with captcha solver and proxy; 3 managed auth connections; 1-day replays; no configurable or custom proxies); Hobbyist $30.00/month ($10 of usage credit per month; 10 concurrent browsers; unlimited proxy usage; 7-day replays; unlimited managed auth connections and vaults); Start-Up $200.00/month ($50 of usage credit per month; 150 concurrent browsers; bring-your-own proxies; GPU acceleration; regional browsers; 30-day replays) - Free tier: Developer plan: $5 of usage credit every month, 5 concurrent browsers, no credit card required. - Isolation: Each browser runs in its own unikernel microVM with its own kernel and filesystem, isolated at the hypervisor - Pause and resume: yes - Snapshots: yes - GPU: yes - Network controls: yes - Max memory GiB: 16 - SDKs: TypeScript, Python, Go, CLI, MCP server, Playwright, CDP - Regions: US by default; regional browsers are available on the Start-Up and Enterprise plans at the same rates. - Open source: yes (Apache-2.0) - Self-hostable: yes - Best for: Agents that wait on users or models between actions, since standby stops charges; Untrusted pages or code next to the browser, given per-browser VM isolation; Proxy-heavy work, since proxy traffic is not metered - Watch out for: Headful browsers cost eight times the headless rate because billing follows memory; Standby browsers still take a concurrency slot until deleted; GPU, custom proxies and regional browsers need the $200/month Start-Up plan; Only 5 concurrent browsers on the free plan - Sources: https://www.kernel.sh/pricing https://www.kernel.sh/docs/info/pricing https://www.kernel.sh/docs/browsers/concurrency-and-limits https://www.kernel.sh/docs/overview/why-kernel https://www.kernel.sh/docs/browsers/standby https://github.com/kernel/kernel-images - Facts checked: 2026-11-13 ## Hyperbrowser Hyperbrowser sells hosted browser sessions for AI agents and scraping, plus sandboxes: isolated Firecracker VMs for running code, tests and agents. Browser time costs $0.10 per hour plus proxy data, and sandboxes cost $0.045 per vCPU-hour and $0.015 per GiB-hour for allocated resources while running. CPU sandboxes can be saved as memory snapshots, including running processes, and started again from them. All usage draws from one credit balance, with monthly plans that include credits. - Page: https://aisandboxhunt.com/hyperbrowser - Company: Hyperbrowser - Website: https://www.hyperbrowser.ai - Categories: Browser sandboxes for AI agents, AI agent sandboxes - Score: 60/100, #4 of 12 browser sandboxes for AI agents (method 1.3 (2026-11-14)) - Per browser-hour: $0.10 - Price per vCPU-hour: $0.045 - Price per GiB-hour of memory: $0.015 - Price per browser-hour: $0.10 - GPU prices per hour: L4 $0.7992, A10G $1.10, L40S $1.95, A100 40GB $2.10, A100 80GB $2.50, RTX PRO 6000 $3.03, H100 $3.95, H200 $4.54, B200 $6.25, B300 $7.10 - Billing increment: Not stated for CPU sandboxes and browsers (rates quoted per hour); GPU time per second - Billed while: Sandboxes: allocated vCPU and memory for the whole time the sandbox runs, including idle time, until stopped or timed out. Browsers: session time plus proxy data. - Plans: Startup $30.00/month (30,000 credits ($30) of usage per month across browsers, sandboxes, web data and agents; 25 concurrent sessions; 30-day data retention; residential proxies); Scale $100.00/month (100,000 credits ($100) of usage per month; 100 concurrent sessions; 30-day data retention; premium residential proxies) - Free tier: Free account with 5,000 credits ($5), enough for about 50 browser-hours; 1 concurrent session; upgrade required once used. - Isolation: Firecracker microVM per CPU sandbox; GPU sandboxes use a gVisor runtime - Snapshots: yes - GPU: yes - Custom images: yes - Network controls: yes - Max vCPU: 32 - Max memory GiB: 256 - SDKs: Node.js, Python, CLI (hx), REST API - Regions: US regions (for example us-west); other regions not stated. - Open source: no - Self-hostable: no - Best for: Agents that need both a hosted browser and a code sandbox under one credit balance; Cheap browser automation at $0.10 per browser-hour before proxies; Restoring prepared environments quickly from memory snapshots - Watch out for: Sandboxes bill allocated vCPU and memory while running, even when idle; Proxy data at $10 per GB can dominate browser costs; GPU and macOS sandboxes need access requests; Free use is a one-time 5,000 credits with a single concurrent session - Sources: https://www.hyperbrowser.ai/pricing https://www.hyperbrowser.ai/docs/pricing https://www.hyperbrowser.ai/docs/api-reference/create-new-sandbox https://www.hyperbrowser.ai/docs/sandboxes/snapshots - Facts checked: 2026-11-13 ## Browserbase Browserbase runs fleets of headless Chromium browsers in the cloud that agents drive with Playwright, Puppeteer, Selenium or its own Stagehand SDK. Each browser runs in a dedicated VM. Plans bundle browser hours, proxy bandwidth, and Search and Fetch API calls; usage past the allowance is billed at $0.12 per browser-hour on Developer and $0.10 on Startup. A free plan includes one browser hour per month with 15-minute sessions. - Page: https://aisandboxhunt.com/browserbase - Company: Browserbase - Website: https://www.browserbase.com - Categories: Browser sandboxes for AI agents - Score: 54/100, #5 of 12 browser sandboxes for AI agents (method 1.3 (2026-11-14)) - Per browser-hour: $0.10 - Price per browser-hour: $0.10 - Billing increment: Per minute, with a one-minute minimum per session - Billed while: While a browser session is running - Plans: Free $0.00/month (1 browser hour per month, 3 concurrent browsers, 15-minute sessions, 5 session creations per minute, no proxies or captcha solving, 3 agent runs, 1,000 Search and 1,000 Fetch calls, $5 of model tokens. No overage); Developer $20.00/month (100 browser hours then $0.12/browser-hour; 1 GB proxy then $12/GB; 25 concurrent browsers; 6-hour sessions; 25 session creations per minute; auto captcha solving; 15 agent runs); Startup $99.00/month (500 browser hours then $0.10/browser-hour; 5 GB proxy then $10/GB; 100 concurrent browsers; 6-hour sessions; 50 session creations per minute; auto captcha solving; 50 agent runs; 10,000 Fetch calls) - Free tier: Free plan: 1 browser hour per month, 3 concurrent browsers and 15-minute sessions, no credit card; usage stops at the allowance. - Isolation: Each browser runs in a dedicated VM, in an isolated subnet - GPU: no - Max session hours: 6 - SDKs: Node.js, Python, Stagehand, Playwright, Puppeteer, Selenium - Regions: us-west-2 (Oregon, default), us-east-1 (Virginia), eu-central-1 (Frankfurt), ap-southeast-1 (Singapore) - Open source: no - Self-hostable: no - Best for: Teams that want Playwright, Puppeteer or Selenium scripts to run against hosted Chromium with session replays; Workloads that fit inside a monthly bundle of 100 or 500 browser hours; Users of the Stagehand SDK who want browsers, search, fetch and model access on one bill - Watch out for: Monthly fee required for anything past 1 browser hour; the free plan stops at its allowance; Proxy traffic is $10-12 per GB beyond the small included amount; Sessions are capped at 6 hours on self-serve plans and 15 minutes on Free; One-minute minimum per session - Sources: https://www.browserbase.com/pricing https://docs.browserbase.com/guides/plans-and-pricing https://docs.browserbase.com/account/enterprise/security https://docs.browserbase.com/optimizations/latency/multi-region - Facts checked: 2026-11-13 ## Anchor Browser Anchor Browser provides hosted Chromium sessions that agents drive over CDP, its SDKs or MCP, with managed authentication, proxies, captcha bypass and a built-in AI task agent. Usage is paid in credits worth $1 each: browser sessions you drive yourself cost $0.01 per browser created plus $0.05 per browser-hour, with proxy data and AI steps extra. Saved automation tasks are instead billed a flat 0.1 credits per successful run, or the infrastructure cost if higher. A free plan gives 5 credits every month. - Page: https://aisandboxhunt.com/anchor-browser - Company: Anchor Browser - Website: https://anchorbrowser.io - Categories: Browser sandboxes for AI agents - Score: 47/100, #6 of 12 browser sandboxes for AI agents (method 1.3 (2026-11-14)) - Per browser-hour: $0.05 - Price per browser-hour: $0.05 - Billing increment: Per minute (active time rounded to the nearest full minute), plus $0.01 per browser created - Billed while: While the browser session is active - Plans: Free $0.00/month (5 credits per month; up to 5 concurrent browsers; up to 500 browser sessions; 5 sessions created per minute; US only; no authenticated browsers, captcha bypass or geolocation. No overage listed); Starter $50.00/month (50 credits per month, then $1.00 per credit; up to 25 concurrent browsers; 25 sessions created per minute; authenticated browsers, captcha bypass, any geolocation; US region); Team $500.00/month (500 credits per month, then $1.00 per credit; up to 50 concurrent browsers; Cloudflare verified browser agents; custom proxy; 30-day data retention); Growth $2,000.00/month (2,000 credits per month, then $1.00 per credit; up to 200 concurrent browsers; 100 sessions created per minute; full stealth browser; US, EU, Asia and Australia regions) - Free tier: Free plan with 5 credits ($5 of usage) every month, up to 5 concurrent browsers. - Isolation: Not stated - GPU: no - SDKs: Python, TypeScript, REST API, MCP, Playwright, Puppeteer - Regions: US on Free and Starter; US, EU, Asia and Australia on Growth; any region on Enterprise. - Open source: no - Self-hostable: no - Best for: Agents that need to sign in to third-party web apps with managed authentication; Long-running browser sessions, since hourly browser time is low and there is no stated session cap; Repeated automation tasks that fit a flat per-run fee - Watch out for: Proxy traffic at $8/GB will usually cost more than browser time; Captcha bypass and geolocation need a paid plan; stealth needs Growth ($2,000/month); Free and Starter plans run in the US only; The pricing page calculator and the docs show different hourly browser rates - Sources: https://docs.anchorbrowser.io/pricing https://anchorbrowser.io/pricing https://docs.anchorbrowser.io/advanced/session-timeout - Facts checked: 2026-11-13 ## Lightpanda Cloud Lightpanda is a headless browser engine built from scratch in Zig with V8 for JavaScript, made for automation, scraping and agents rather than for display. Lightpanda Cloud runs it for you behind a CDP endpoint that Puppeteer, Playwright or chromedp connect to, plus an HTTP API and an MCP server, in west Europe and west US. The free Explorer plan includes 10 browser hours a month; Builder is $19 a month for 300 hours, then $0.08 per hour. The browser itself is AGPL-3.0 open source and can be run locally. - Page: https://aisandboxhunt.com/lightpanda - Company: Lightpanda - Website: https://lightpanda.io - Categories: Browser sandboxes for AI agents - Score: 43/100, #8 of 12 browser sandboxes for AI agents (method 1.3 (2026-11-14)) - Per browser-hour: $0.08 - Price per browser-hour: $0.08 - Billing increment: Not stated - Billed while: Browser hours; how session time is measured is not stated - Plans: Explorer $0.00/month (10 browser hours per month; 5 concurrent sessions; bring your own proxy; community support on Discord. Overage pricing is not stated); Builder $19.00/month (300 browser hours per month, then $0.08 per hour; 30 concurrent sessions; bring your own proxy; email and Slack support) - Free tier: Explorer plan: 10 browser hours every month, free with no end date, 5 concurrent sessions. - Isolation: Not stated - GPU: no - SDKs: CDP (Puppeteer, Playwright, chromedp), HTTP API, MCP server - Regions: West Europe (euwest) and west US (uswest). - Open source: yes (AGPL-3.0) - Self-hostable: yes - Best for: High-volume scraping and agent page reads over CDP where a lighter, faster browser helps; Teams that bring their own proxies; Running the same engine locally for free and in the cloud for scale - Watch out for: It is not Chromium, so some sites or scripts may behave differently; No visual rendering, so screenshots and pixel-based agents are not supported; Only two regions; Billing increment and session time limits are not published - Sources: https://lightpanda.io/pricing https://lightpanda.io/docs/cloud-offer/getting-started https://lightpanda.io/llms.txt https://github.com/lightpanda-io/browser - Facts checked: 2026-11-13 ## Browser Use Cloud Browser Use Cloud hosts stealth Chromium browsers that you can drive over CDP with Playwright or Puppeteer, or hand tasks to its hosted web agent. Browsers cost $0.06 per hour, billed by the minute with a one-minute minimum; traffic is billed separately at $5/GB through home-IP proxies or $0.20/GB direct, and CAPTCHA solving is included. There is no subscription: you buy prepaid credits that do not expire, and the number of browsers you can run at once rises with your total spend. New accounts get a one-time $1 credit. - Page: https://aisandboxhunt.com/browser-use - Company: Browser Use - Website: https://browser-use.com - Categories: Browser sandboxes for AI agents - Score: 42/100, #9 of 12 browser sandboxes for AI agents (method 1.3 (2026-11-14)) - Per browser-hour: $0.06 - Price per browser-hour: $0.06 - Billing increment: Per minute, rounded up, with a one-minute minimum - Billed while: While the browser session runs; the requested session time is charged upfront and unused time is refunded when the session is stopped - Plans: Pay as you go $0.00/month (Prepaid credits from $5 (1 credit = $1) that never expire. 10 concurrent browsers to start; 50 after $100 total paid, 250 after $1,000, 500 after $5,000, 1,000 after $25,000. Sessions up to 4 hours) - Free tier: One-time $1 signup credit, one per person. - Isolation: Not stated - GPU: no - Max session hours: 4 - SDKs: Python, TypeScript, REST API, Playwright, Puppeteer - Regions: Not stated for self-serve; EU data residency is offered on Enterprise. - Open source: no - Self-hostable: no - Best for: High-volume browser work at a low hourly rate without a subscription; Bursty workloads, since credits never expire; Teams already using the open-source Browser Use agent library - Watch out for: Traffic is metered even without a proxy ($0.20/GB), and home-IP proxy traffic at $5/GB usually costs more than browser time; Only 10 concurrent browsers until you have paid $100 in total; Sessions are limited to 4 hours; The free credit is only $1 and is one-time - Sources: https://browser-use.com/pricing https://docs.browser-use.com/cloud/api-v2/browsers/create-browser-session - Facts checked: 2026-11-13 ## Steel Steel runs cloud browser sessions that agents control over CDP with Puppeteer, Playwright or Selenium, or through its own SDKs, CLI and MCP server. Browser time is billed by the minute at $0.10 per hour on the free Launch plan and $0.08 on the $250/month Scale plan, with proxy bandwidth and CAPTCHA solves charged separately. New accounts get $30 of one-time credit. The underlying browser server, Steel Browser, is Apache-2.0 licensed and can be run with Docker. - Page: https://aisandboxhunt.com/steel - Company: Nen Labs, Inc. - Website: https://steel.dev - Categories: Browser sandboxes for AI agents - Score: 41/100, #10 of 12 browser sandboxes for AI agents (method 1.3 (2026-11-14)) - Per browser-hour: $0.08 - Price per browser-hour: $0.08 - Billing increment: Per minute, rounded up - Billed while: While a browser session is live, until it is released or times out - Plans: Launch $0.00/month ($0.10/browser-hour, $10/GB proxy, $3 per 1,000 CAPTCHA solves; 10 concurrent sessions; sessions up to 15 minutes; 60 requests per minute; up to 3 seats); Scale $250.00/month ($100 of usage credit every month; $0.08/browser-hour, $6/GB proxy, $1 per 1,000 CAPTCHA solves; 100 concurrent sessions; sessions up to 1 hour; 600 requests per minute; dedicated IPs at $5 each) - Free tier: $30 of one-time credit on the Launch plan, no credit card needed. - Isolation: Each session is an isolated cloud browser with its own filesystem; the docs mention a session VM but do not name the technology - GPU: no - Max session hours: 1 - SDKs: TypeScript/Node.js, Python, CLI, REST API, MCP server - Regions: Hosted browsers run in one region, us-east; proxy geolocation sets the public IP location. - Open source: yes (Apache-2.0) - Self-hostable: yes - Best for: Short agent or scraping sessions billed by the minute without a monthly fee; Teams that want the option to run the same browser server themselves; Agents driven from a CLI or MCP client as well as from code - Watch out for: Sessions on the free plan end after 15 minutes; 1 hour is the self-serve maximum; The $30 starting credit is one-time, not monthly; Hosted browsers run only in us-east; Stealth browser is only on Enterprise - Sources: https://steel.dev/pricing https://docs.steel.dev/overview/pricinglimits https://docs.steel.dev/overview/sessions-api/configuration https://github.com/steel-dev/steel-browser - Facts checked: 2026-11-13 ## Airtop Airtop runs cloud browser sessions that agents control through its API with natural-language actions such as page queries and form filling, or directly with Puppeteer, Playwright or Selenium. It also sells Agent Builder, which compiles described workflows into scheduled agents, and Mark, a marketing assistant. All usage, including browser time, proxy bandwidth and AI calls, is paid in credits from a monthly plan, but Airtop does not publish how many credits a browser-hour uses. Plans run from a free tier with 1,000 credits a month to $558 a month. - Page: https://aisandboxhunt.com/airtop - Company: Airtop - Website: https://www.airtop.ai - Categories: Browser sandboxes for AI agents - Score: 38/100, #11 of 12 browser sandboxes for AI agents (method 1.3 (2026-11-14)) - Per browser-hour: No public rate - Billing increment: Not stated - Billed while: Credits are charged for the browser's CPU time, proxy bandwidth and AI calls; sessions end after 10 minutes of inactivity by default - Plans: Free $0.00/month (1,000 credits per month; 3 simultaneous sessions; 1 deployed agent. No credit purchases beyond the plan); Starter $29.00/month (30,000 credits per month at the base price (larger allowances up to 150,000 credits available at higher prices); 3 simultaneous sessions; 10 deployed agents; integrated proxy); Professional $189.00/month (225,000 credits per month at the base price (up to 500,000 at higher prices); 30 simultaneous sessions; 30 deployed agents; custom proxy); Enterprise $558.00/month (775,000 credits per month at the base price (up to 1,500,000 at higher prices); 100 simultaneous sessions; unlimited deployed agents; SOC 2 Type 2 report) - Free tier: Free plan with 1,000 credits every month and 3 simultaneous sessions, plus a one-time signup credit bonus in the first month. - Isolation: Each session runs in its own encrypted environment; the isolation technology is not stated - GPU: no - SDKs: Node.js, Python, REST API, Puppeteer, Playwright, Selenium, n8n, Make, Zapier - Regions: Not stated - Open source: no - Self-hostable: no - Best for: Agents that want natural-language page queries and actions instead of writing selectors; No-code automation builders using n8n, Make or Zapier; Workflows behind logins using saved browser profiles - Watch out for: No published price per browser-hour, so costs are hard to compare or predict; No pay-as-you-go overage: when credits run out you must upgrade the plan; Only 3 simultaneous sessions on Free and Starter; The company signals a product change, and its focus has moved toward agent building - Sources: https://www.airtop.ai/pricing https://www.airtop.ai/help/billing/billing https://docs.airtop.ai/guides/misc/faq https://docs.airtop.ai/guides/how-to/creating-a-session https://www.airtop.ai/web-automation - Facts checked: 2026-11-13 ## Browserless Browserless hosts browsers that existing Puppeteer or Playwright scripts connect to over WebSocket, plus BrowserQL, REST APIs for screenshots, PDFs and scraping, and a managed agent API. Plans include a monthly number of units, where one unit is up to 30 seconds of browser time per connection; extra units cost $0.0015 to $0.0020 each, or $0.18 to $0.24 per browser-hour. Residential and datacenter proxies and captcha solves are also paid in units. The source is available under the SSPL or a commercial license and can be self-hosted with Docker. - Page: https://aisandboxhunt.com/browserless - Company: Browserless - Website: https://www.browserless.io - Categories: Browser sandboxes for AI agents - Score: 35/100, #12 of 12 browser sandboxes for AI agents (method 1.3 (2026-11-14)) - Per browser-hour: $0.18 - Price per browser-hour: $0.18 - Billing increment: Units of up to 30 seconds per browser connection; each reconnect starts a new unit - Billed while: While a browser connection is open - Plans: Free $0.00/month (1,000 units per month (up to about 8.3 browser-hours); 2 concurrent browsers; 2-minute sessions; residential and datacenter proxies and captcha solving paid from units; no overage); Prototyping $35.00/month (20,000 units per month, then $0.0020 per unit; 5 concurrent browsers (10 if billed annually at $25/month); 15-minute sessions; bring your own proxy; screen recording); Starter $200.00/month (180,000 units per month, then $0.0017 per unit; 30 concurrent browsers (40 if billed annually at $140/month); 30-minute sessions); Scale $500.00/month (500,000 units per month, then $0.0015 per unit; 80 concurrent browsers (100 if billed annually at $350/month); 60-minute sessions) - Free tier: Free plan: 1,000 units (30-second blocks) every month, 2 concurrent browsers, 2-minute sessions, no card required. - Isolation: Not stated - GPU: no - Max session hours: 1 - SDKs: TypeScript, Python, BrowserQL, REST API, Puppeteer, Playwright, MCP - Regions: San Francisco, London and Amsterdam endpoints; more locations, including the EU, on Enterprise. - Open source: no - Self-hostable: yes - Best for: Existing Puppeteer or Playwright scripts that only need a new connection URL; Screenshot, PDF and scraping jobs through ready-made REST endpoints; Teams that may later move to a self-hosted deployment - Watch out for: Billed per 30-second unit per connection, so many short connections round up; Proxy traffic is paid in units and can use them up quickly; Sessions are capped at 60 minutes even on the largest self-serve plan, and 2 minutes on Free; Commercial self-hosting needs a paid license - Sources: https://www.browserless.io/pricing https://docs.browserless.io/llms.txt https://github.com/browserless/browserless/blob/main/LICENSE - Facts checked: 2026-11-13 # GPU sandboxes Sandbox APIs that can attach a GPU. Compared on the hourly price of one NVIDIA H100, the most common shared model. Ranking: https://aisandboxhunt.com/gpu-sandboxes ## Thunder Compute Thunder Compute rents GPUs two ways. Instances are persistent GPU machines you create from a console, CLI or VS Code, billed per minute, from $0.35/hour for an RTX A6000 to $3.20/hour for an H100. Sandboxes are Firecracker microVMs for agents, RL and evals, created from a Python SDK, MCP server or REST API and billed per second for each GPU, vCPU, memory and storage. Sandbox access must be enabled for your organization on request, and there are no egress fees. - Page: https://aisandboxhunt.com/thunder-compute - Company: Thunder Compute - Website: https://www.thundercompute.com - Categories: GPU sandboxes - Score: 50/100, #6 of 7 GPU sandboxes (method 1.3 (2026-11-14)) - H100 per hour: $2.95 - Price per vCPU-hour: $0.051 - Price per GiB-hour of memory: $0.016 - GPU prices per hour: H100 SXM (sandbox) $2.95, A100 80GB (sandbox) $1.95, H100 SXM (instance, 4 vCPU included) $3.20, A100 80GB (instance, 8 vCPU included) $1.09, L40 (instance, 6 vCPU included) $0.79, RTX A6000 (instance, 6 vCPU included) $0.35 - Storage per GiB-month: $0.0876 - Billing increment: per second (sandboxes); per minute (instances) - Billed while: Sandboxes: each resource per second until the sandbox stops or its TTL ends. Instances: while running; snapshots bill until deleted. - Free tier: No free tier or trial credit is stated on the pricing or billing pages. - Isolation: Sandboxes: Firecracker microVM per sandbox. Instances: EC2-based machines sharing a pool of network-attached GPUs - Snapshots: yes - GPU: yes - Custom images: yes - Network controls: yes - SDKs: Python (thunder-sandbox), CLI (tnr), REST API, MCP server, VS Code extension - Regions: Not stated - Open source: no - Self-hostable: no - Best for: Low hourly GPU prices for development, notebooks and training on persistent instances; GPU-backed RL rollouts, evals and CI from code, once sandbox access is enabled; Workloads that move a lot of data out, since egress is free - Watch out for: Sandbox access requires a request; only instances are fully self-serve; Instances share a pool of network-attached GPUs rather than reserving one; No free tier or trial credit is published; Regions are not stated - Sources: https://www.thundercompute.com/pricing https://www.thundercompute.com/docs/sandboxes/quickstart https://www.thundercompute.com/docs/faq/faq https://www.thundercompute.com/products/sandboxes - Facts checked: 2026-11-13 # CI runners Hosted build and test runners, mostly for GitHub Actions. Compared on a 2 vCPU Linux runner. Ranking: https://aisandboxhunt.com/ci-runners ## Ubicloud GitHub Actions runners Ubicloud runs GitHub Actions jobs in a fresh, ephemeral VM per job on its own cloud built on Cloud Hypervisor and bare-metal hosts in Germany. You connect a GitHub App and change the runs-on label; x64 and arm64 runners come in 2 to 30 vCPU sizes. Standard runners cost $0.00125 per minute for 2 vCPU and premium runners, enabled by default, $0.002, both billed by the minute. Every account gets a $2.5 monthly credit, and the platform source code is available under AGPL-3.0. - Page: https://aisandboxhunt.com/ubicloud - Company: Ubicloud - Website: https://www.ubicloud.com - Categories: CI runners - Score: 79/100, #1 of 10 CI runners (method 1.3 (2026-11-14)) - List price per month: $7.50 for 2 vCPU Linux runner, 100 hours a month - Fixed-size price per hour: $0.075 (Standard x64 runner: 2 vCPU, 8 GB, Linux ($0.00125 per minute)) - Billing increment: per minute - Billed while: While a job's runner VM is provisioned; the VM is removed when the job ends. - Plans: Pay as you go $0.00/month ($2.5 monthly account credit (1,250 standard 2 vCPU runner minutes), then per-minute rates. A valid credit card is required before creating resources) - Free tier: $2.5 credit every month, equal to 1,250 minutes on the 2 vCPU standard runner. A credit card is required to start. - Isolation: Clean, ephemeral VM per job on Cloud Hypervisor - Pause and resume: no - Max vCPU: 30 - Max memory GiB: 120 - SDKs: GitHub Actions runs-on labels, GitHub App - Regions: Managed runners are placed across data centers in Germany. - Open source: yes (AGPL-3.0) - Self-hostable: yes - Best for: Teams that want the lowest per-minute price for Linux GitHub Actions jobs; Workloads that need arm64 runners at the same price as x64; Teams that prefer an open-source platform they could run themselves - Watch out for: Runners run only in Germany; Linux only; no macOS or Windows runners; Premium runners are on by default and cost 60% more than standard; A credit card is required even to use the free monthly credit - Sources: https://www.ubicloud.com/docs/about/pricing https://www.ubicloud.com/docs/github-actions-integration/runner-types https://www.ubicloud.com/docs/github-actions-integration/security https://www.ubicloud.com/docs/github-actions-integration/quickstart https://www.ubicloud.com/docs/github-actions-integration/ubicloud-cache https://github.com/ubicloud/ubicloud - Facts checked: 2026-11-13 ## RunsOn RunsOn is a control plane you deploy into your own AWS account with one CloudFormation or Terraform stack; it launches a fresh EC2 instance for every GitHub Actions job and removes it afterwards. You pay RunsOn a flat yearly license sized by jobs per month (from $350 a year), and AWS bills you directly for the instances, disks, cache bucket and a small control plane. On its pricing page RunsOn estimates a 2 vCPU x64 spot runner at about $0.0009 a minute in us-east-1. It is free for open-source, non-profit and personal non-commercial use. - Page: https://aisandboxhunt.com/runs-on - Company: RunsOn - Website: https://runs-on.com - Categories: CI runners - Score: 75/100, #2 of 10 CI runners (method 1.3 (2026-11-14)) - List price per month: $5.40 for 2 vCPU Linux runner, 100 hours a month - Fixed-size price per hour: $0.054 (AWS m8a.large spot, 2 vCPU x64, with a 30 GB gp3 root volume; paid to AWS, not RunsOn ($0.0009 per minute, us-east-1 7-day average)) - GPU prices per hour: NVIDIA T4 (AWS g4dn.xlarge spot, 4 vCPU; paid to AWS) $0.234 - Billing increment: per second (AWS EC2 and EBS billing); license billed yearly - Billed while: AWS bills the instance and its volume only while a job runs; the control plane (about $7 a month) runs continuously in your account. The license is a flat yearly fee. - Plans: Free (non-commercial) $0.00/month (Full product for open-source, non-profit and personal non-commercial projects, in exchange for a public link back to RunsOn. AWS costs still apply); Starter $29.17/month ($350 a year (€300) for under 50,000 jobs a month; all GitHub orgs and AWS accounts of one company; email support. AWS compute billed separately); Growth $87.50/month ($1,050 a year (€900) for under 200,000 jobs a month. AWS compute billed separately); Scale $175.00/month ($2,100 a year (€1,800) for under 500,000 jobs a month. AWS compute billed separately); Sponsorship $350.00/month ($4,200 a year (€3,600), unlimited jobs, full server and agent source code for review, priority support via Slack Connect. Also sold on AWS Marketplace. AWS compute billed separately) - Free tier: 15-day free trial of Starter, Growth or Scale (payment details collected at checkout); AWS usage during the trial is billed by AWS. Open-source, non-profit and personal non-commercial projects get a free license. - Isolation: Fresh EC2 instance per job in your own AWS account - Pause and resume: no - GPU: yes - Custom images: yes - Max vCPU: 96 - SDKs: GitHub Actions runs-on labels, CloudFormation, Terraform - Regions: Any AWS account and region where you deploy the stack; one license covers all your stacks. Example prices are for us-east-1. - Open source: no - Self-hostable: yes - Best for: Companies with heavy GitHub Actions usage that want compute at AWS spot prices; Teams that need runners inside their own AWS network, IAM and VPC; Jobs that need specific EC2 families, GPUs or custom AMIs - Watch out for: You operate the stack in your own AWS account and pay AWS separately; Spot prices vary and spot instances can be reclaimed mid-job; The yearly license is due even in months with little CI use; AWS only; no macOS runners - Sources: https://runs-on.com/pricing/ https://github.com/runs-on/runs-on - Facts checked: 2026-11-13 ## Bitrise CI Bitrise is a CI/CD platform built for iOS and Android teams, running builds on managed macOS (Apple silicon) and Ubuntu virtual machines with a library of ready-made workflow steps. The free Hobby plan includes 500 build minutes a month; paid Starter and Pro plans are priced per month by concurrency and builds rather than by minute, and each machine type has a published per-minute cost used for plan minutes and pay-as-you-go. The smallest Linux machine, Linux 2S (2 vCPU, 8 GB), costs $0.0022 a minute and is available on Pro. - Page: https://aisandboxhunt.com/bitrise - Company: Bitrise - Website: https://bitrise.io - Categories: CI runners - Score: 74/100, #3 of 10 CI runners (method 1.3 (2026-11-14)) - List price per month: $13.20 for 2 vCPU Linux runner, 100 hours a month - Fixed-size price per hour: $0.132 (Linux 2S: Ubuntu VM, 2 vCPU, 8 GB ($0.0022 per minute; Pro plan)) - Billing increment: per minute - Billed while: While a build runs; minutes count against the plan's allowance, then pay-as-you-go where the plan allows it. - Plans: Hobby $0.00/month (500 build minutes a month on Linux or macOS (Linux Medium 4 vCPU / 16 GB or M2 Pro Medium), 1 private app, 90-minute build timeout. Optional pay-as-you-go at 1.5x the per-minute rate, capped at $15 a month); Starter $99.00/month (Unlimited usage with 1 concurrent build on Tier 1 ($99/month; $199 for 2 and $299 for 4 concurrent builds). Up to 3 private apps, Linux M (4 vCPU / 16 GB) and macOS Medium machines, 210-minute build timeout. $89/month billed annually); Pro $218.00/month (From $218/month ($200 billed annually) for 250 builds a month; tiers include 3,000 to 24,000 zero-margin infrastructure minutes. Up to 10 concurrent macOS and 30 Linux builds; Linux 2S, Linux L (8 vCPU / 32 GB) and M2 Pro Large machines. Pay-as-you-go and extra build capacity available) - Free tier: Hobby plan: free forever with 500 build minutes a month on Linux or macOS. New sign-ups also get a 30-day trial of the paid plans. - Isolation: Managed Ubuntu or macOS virtual machine per build - Pause and resume: no - Max session hours: 3.5 - Max vCPU: 8 - Max memory GiB: 32 - SDKs: Workflow YAML, MCP server - Regions: Not stated on the pricing page; Enterprise can use dedicated environments or self-hosted runners on AWS. - Open source: no - Self-hostable: yes - Best for: iOS and Android teams that need managed macOS and Xcode builds; Small mobile projects that fit in 500 free minutes a month; Teams that prefer a fixed monthly price per concurrent build over per-minute billing - Watch out for: Paid plans are priced by builds and concurrency, so per-minute costs are hard to compare; The cheapest 2 vCPU Linux machine needs the Pro plan, which starts at $218 a month; Builds time out after 210 minutes on self-serve plans (90 on Hobby); Pro pay-as-you-go rates are not published - Sources: https://bitrise.io/pricing - Facts checked: 2026-11-14 ## GitHub-hosted runners GitHub-hosted runners run each GitHub Actions job on a new virtual machine (single-CPU runners use a container on a shared VM), with Linux, Windows and macOS images maintained by GitHub. In private repositories usage is priced per minute, rounded up to the whole minute, and drawn first from a monthly dollar allowance included with every GitHub plan. Standard runners are free in public repositories. Larger runners up to 96 cores, GPU runners and custom images are available on Team and Enterprise Cloud. - Page: https://aisandboxhunt.com/github-actions-runners - Company: GitHub - Website: https://github.com/features/actions - Categories: CI runners - Score: 70/100, #5 of 10 CI runners (method 1.3 (2026-11-14)) - List price per month: $21.00 for 2 vCPU Linux runner, 100 hours a month - Fixed-size price per hour: $0.36 (Standard Linux 2-core x64 runner for private repositories: 2 vCPU, 8 GB ($0.006 per minute)) - GPU prices per hour: Linux 4-core GPU runner (GPU model not stated) $3.12 - Storage per GiB-month: $0.07 - Billing increment: per minute; each job is rounded up to the nearest whole minute - Billed while: Only while a workflow job runs on the runner; standard runners are free in public repositories. - Plans: GitHub Free $0.00/month ($12 of included compute a month (2,000 minutes on standard Linux), 500 MB artifact storage, 10 GB cache per repository. Without a payment method, jobs stop once the allowance is used. Larger runners are not available); GitHub Team $4.00/month ($4 per user a month; $18 of included compute a month (3,000 minutes on standard Linux), 2 GB artifact storage, 75 GB custom image storage. Larger runners available and can draw from the allowance); GitHub Enterprise Cloud $21.00/month ($21 per user a month; $300 of included compute a month (50,000 minutes on standard Linux), 50 GB artifact storage, 150 GB custom image storage) - Free tier: Standard runners are free and unlimited in public repositories. For private repositories every plan includes a monthly compute allowance: $12 on GitHub Free (2,000 standard Linux minutes). - Isolation: New virtual machine per job (single-CPU runners share a VM as containers) - Pause and resume: no - Snapshots: yes - GPU: yes - Custom images: yes - Max session hours: 6 - Max vCPU: 96 - SDKs: GitHub Actions workflows, REST API, GitHub CLI - Regions: Linux jobs run on a multi-cloud pool across AWS and Microsoft Azure; Windows and macOS runners run in Azure. Region choice for private networking is available on larger runners. - Open source: no - Self-hostable: no - Best for: Public open-source repositories, where standard runners are free; Small private projects that fit inside the monthly included allowance; Teams that want Linux, Windows, macOS and GPU runners without another vendor - Watch out for: Per-minute rates are higher than most third-party runner services; Every job is rounded up to a whole minute; Jobs are capped at 6 hours; Larger and GPU runners need a paid organization plan and are billed in public repositories too - Sources: https://docs.github.com/en/billing/reference/actions-runner-pricing https://docs.github.com/en/billing/concepts/product-billing/github-actions https://github.com/pricing https://docs.github.com/en/actions/concepts/runners/github-hosted-runners https://docs.github.com/en/actions/reference/limits - Facts checked: 2026-11-13 ## WarpBuild WarpBuild provides GitHub Actions runners on ephemeral VMs that are created for each job and destroyed afterwards. Hosted cloud runners cover Linux x64 and arm64 (2 to 32 vCPU), Windows and macOS on Apple M4/M5 Pro, priced per minute with no base fee; a 2 vCPU x64 runner is $0.004 a minute. A bring-your-own-cloud option runs runners in your AWS, GCP or Azure account for a $0.002 per-minute fee plus your cloud bill. Add-ons include cache storage, snapshot runners and remote Docker builders. - Page: https://aisandboxhunt.com/warpbuild - Company: WarpBuild - Website: https://www.warpbuild.com - Categories: CI runners - Score: 66/100, #7 of 10 CI runners (method 1.3 (2026-11-14)) - List price per month: $24.00 for 2 vCPU Linux runner, 100 hours a month - Fixed-size price per hour: $0.24 (Linux x86-64 cloud runner: 2 vCPU, 8 GB ($0.004 per minute)) - Storage per GiB-month: $0.20 - Billing increment: per minute - Billed while: While a job runs on the runner VM; the VM and its disk are deleted when the job ends. Cache and snapshot storage are billed separately. - Plans: Cloud runners $0.00/month (Usage-based with no base, platform or seat fee. Unlimited job concurrency. Add-ons: cache storage $0.20/GB-month, cache operations $0.0001 each, snapshot restore $0.04 per job, snapshot storage $0.025 per hour per snapshot; SAML SSO and data residency $250/month each, audit logs $400/month); Bring your own cloud $0.00/month (Runners in your own AWS, GCP or Azure account for $0.002 per minute (Linux, and Windows on AWS and Azure), plus the compute bill from your cloud provider. Cache and snapshots included) - Free tier: One-time $10 credit on signup, no credit card required; a payment method is needed once the free credit is used. - Isolation: Ephemeral VM per job, destroyed when the workflow completes - Pause and resume: no - Snapshots: yes - GPU: no - Max vCPU: 32 - Max memory GiB: 128 - SDKs: GitHub Actions runs-on labels, REST API - Regions: Not stated for the hosted cloud; enterprise plans can pin runners to a chosen region, and the bring-your-own-cloud option runs in your own AWS, GCP or Azure region. - Open source: no - Self-hostable: yes - Best for: Teams that want cheaper GitHub Actions minutes with no base fee; Organizations that want runners inside their own AWS, GCP or Azure account; Jobs with long setup that can resume from a saved runner snapshot - Watch out for: Smallest Windows runner is 4 vCPU at $0.016 per minute; Cache storage, cache operations and snapshots are billed on top of minutes; Hosted cloud regions are not published; Free usage is a one-time $10 credit, not a monthly allowance - Sources: https://www.warpbuild.com/pricing https://www.warpbuild.com/answers/how-much-do-github-actions-runners-cost https://www.warpbuild.com/docs/ci/cloud-runners - Facts checked: 2026-11-13 ## Blacksmith Blacksmith runs GitHub Actions jobs in ephemeral Firecracker microVMs on its own bare-metal fleet in the US and EU, switched in by changing the runs-on label. Linux x64 and ARM, Windows and macOS M4 runners are available from 2 to 32 vCPU, with a co-located cache, Docker layer caching and sticky disks. Ubuntu x64 costs $0.004 per minute for 2 vCPU and scales linearly with size. Every organization gets 3,000 free minutes a month, and there is no concurrency cap. - Page: https://aisandboxhunt.com/blacksmith - Company: Blacksmith Software Inc - Website: https://www.blacksmith.sh - Categories: CI runners - Score: 62/100, #8 of 10 CI runners (method 1.3 (2026-11-14)) - List price per month: $24.00 for 2 vCPU Linux runner, 100 hours a month - Fixed-size price per hour: $0.24 (Ubuntu x64 runner: 2 vCPU, 8 GB, Linux ($0.004 per minute)) - Storage per GiB-month: $0.50 - Billing increment: per minute - Billed while: While a job runs on the runner; the microVM is destroyed when the job ends. Sticky disks and Docker layer cache are billed per GB-month. - Plans: Pay as you go $0.00/month (3,000 free minutes a month, then per-minute rates; no concurrency limit. Add-ons: Docker layer caching and sticky disks $0.50/GB-month, static IPs $100/IP-month) - Free tier: 3,000 free minutes every month; no credit card required to start. The pricing page does not say which runner size the free minutes are counted on. - Isolation: Ephemeral Firecracker microVM per job on Blacksmith-run bare metal - Pause and resume: no - Max vCPU: 32 - Max memory GiB: 128 - SDKs: GitHub Actions runs-on labels, CLI - Regions: us-east (Ashburn), us-west (Phoenix), eu-west (Amsterdam) and eu-central (Frankfurt); moving an organization to another region is done through support. - Open source: no - Self-hostable: no - Best for: Teams that want faster single-thread CPUs for GitHub Actions with a one-line change; Docker-heavy pipelines that benefit from layer caching and sticky disks; Bursty workloads that shard tests widely, since concurrency is not capped - Watch out for: Caching add-ons are billed per GB-month on top of minutes; Static IPs cost $100 per IP per month; macOS runners cost 20 times the 2 vCPU Linux rate per minute; Only works with GitHub Actions - Sources: https://www.blacksmith.sh/pricing https://docs.blacksmith.sh/blacksmith-runners/overview https://www.blacksmith.sh/security https://docs.blacksmith.sh/introduction/quickstart - Facts checked: 2026-11-13 ## Depot Depot sells fast CI: its own Depot CI, drop-in GitHub Actions runners and remote container builds. The same compute also powers Depot Sandboxes, Linux microVMs created from a Node.js SDK, and Depot Managed Agents. CI sandboxes and GitHub Actions runners cost $0.0001 per second for 2 vCPU and 8 GB, scaling linearly to 64 vCPU and 256 GB. Plans start at $20 a month with included minutes. - Page: https://aisandboxhunt.com/depot - Company: Depot - Website: https://depot.dev - Categories: CI runners, AI agent sandboxes - Score: 58/100, #9 of 10 CI runners (method 1.3 (2026-11-14)) - List price per month: $56.00 for 2 vCPU Linux runner, 100 hours a month - Fixed-size price per hour: $0.36 (2 vCPU / 8 GB (base CI sandbox or GitHub Actions runner)) - Billing increment: per second (Depot CI and Sandboxes); per minute (GitHub Actions runners) - Billed while: While the job or sandbox runs, by allocated vCPU; memory is included in the vCPU price. - Plans: Developer $20.00/month (1 user. Each month: 2,000 sandbox minutes (shared by Depot CI, Sandboxes and Managed Agents, measured at 2 vCPUs), 2,000 GitHub Actions minutes, 500 container build minutes, 25 GB cache and registry storage); Startup $200.00/month (Unlimited users. Each month: 20,000 sandbox minutes (shared, measured at 2 vCPUs), 20,000 GitHub Actions minutes, 5,000 container build minutes, 250 GB cache and registry storage) - Free tier: 7-day free trial, no credit card required. - Isolation: Depot CI and Sandboxes run in microVMs on Depot Metal; GitHub Actions runners use ephemeral single-tenant EC2 instances - Snapshots: yes - GPU: no - Custom images: yes - Network controls: yes - Max session hours: 24 - Max vCPU: 64 - Max memory GiB: 256 - SDKs: Node.js (@depot/sandbox), CLI, API, GitHub Actions runner labels - Regions: US and EU; custom regions on the Business plan. - Open source: no - Self-hostable: no - Best for: Speeding up GitHub Actions with a one-line runner label change; Teams that want CI and agent sandboxes on one bill and one allowance; Large CI jobs that need up to 64 vCPU and 256 GB - Watch out for: No free plan after the 7-day trial; paid plans start at $20 a month; Sandbox SDK is Node.js only and custom images must be Depot snapshot images; Sandboxes live at most 24 hours, and snapshots keep files only; Billed on allocated vCPU for the whole run, not on CPU actually used - Sources: https://depot.dev/pricing https://depot.dev/docs/ci/overview https://depot.dev/docs/github-actions/runner-types https://depot.dev/docs/sandboxes/overview https://depot.dev/docs/api/sandbox-sdk-reference https://depot.dev/docs/github-actions/overview https://depot.dev/docs - Facts checked: 2026-11-13 ## Buildkite hosted agents Buildkite hosted agents run Buildkite Pipelines jobs on Buildkite-managed machines, with each job in an isolated virtualized environment that is destroyed when the job finishes. Linux agents come in Small (2 vCPU, 4 GB), Medium and Large shapes on AMD64 and ARM64, with Mac M4 and Windows agents on Pro. Linux usage costs $0.004 per vCPU-minute, metered to the second with no minimums, and cache volumes add no charge. The Free plan includes 2,000 Linux vCPU-minutes a month; Pro is $30 per active user and includes 4,000. - Page: https://aisandboxhunt.com/buildkite - Company: Buildkite - Website: https://buildkite.com - Categories: CI runners - Score: 53/100, #10 of 10 CI runners (method 1.3 (2026-11-14)) - List price per month: $78.00 for 2 vCPU Linux runner, 100 hours a month - Price per vCPU-hour: $0.24 - Fixed-size price per hour: $0.48 (Linux Small: 2 vCPU, 4 GB, AMD64 or ARM64 ($0.008 per minute)) - Billing increment: per second, in vCPU-minutes, no minimum - Billed while: Only while a job runs; startup and shutdown are excluded. Cache volumes are not charged. - Plans: Free $0.00/month (Up to 2,000 Linux AMD64 vCPU-minutes a month on the Small shape only, up to 20 Linux vCPUs at once, up to 10 concurrent jobs, 5 users. No ARM64, Mac or Windows hosted agents); Pro $30.00/month ($30 per active user a month, up to 50 users. Includes 4,000 Linux AMD64 vCPU-minutes a month, then per-minute rates. Up to 256 Linux AMD64 vCPUs, 64 ARM64, 48 Mac and 64 Windows vCPUs at once; Small, Medium and Large shapes. 10 self-hosted agents included, then $3.50 per agent a month) - Free tier: Free plan with up to 2,000 Linux vCPU-minutes a month (about 16.7 hours on the 2 vCPU Small shape). New organizations also get a 30-day all-access trial with no credit card. - Isolation: Ephemeral virtualized environment per job with hypervisor-level isolation between clusters; Linux jobs run in a container image inside it - Pause and resume: no - Custom images: yes - Max session hours: 8 - Max vCPU: 8 - Max memory GiB: 32 - SDKs: Pipeline YAML, Buildkite agent, REST and GraphQL APIs - Regions: Hosted agents run only in US East Coast data centers. - Open source: no - Self-hostable: yes - Best for: Teams already on Buildkite Pipelines that want to stop running their own agents; Mixed fleets that combine hosted agents with self-hosted agents in one pipeline; Short jobs, since billing is metered to the second with no minimum - Watch out for: Pro is charged per active user on top of compute; The Free plan offers only the 2 vCPU / 4 GB Linux shape; Hosted agents run only in US East Coast data centers; Only works with Buildkite Pipelines, not GitHub Actions - Sources: https://buildkite.com/pricing https://buildkite.com/docs/agent/buildkite-hosted https://buildkite.com/docs/agent/buildkite-hosted/linux - Facts checked: 2026-11-13 # Coding agents with cloud machines Coding agents that run tasks on machines the vendor hosts. Most include the machine in a plan price, so we compare features and limits, not hourly cost. Ranking: https://aisandboxhunt.com/coding-agents ## Claude Code on the web Claude Code cloud sessions run in an isolated Ubuntu VM that Anthropic manages, started from claude.ai/code, the Claude apps, the desktop app or the terminal with claude --cloud. Each VM has about 4 vCPUs, 16 GB of RAM and 30 GB of disk, pauses when idle with its files saved, and starts from a cached snapshot of your setup script. There is no separate charge for the VM: sessions count toward the plan's shared Claude usage limits, available from the $20 Pro plan upward. Network access is limited by default and can be widened or turned off per environment. - Page: https://aisandboxhunt.com/claude-code-web - Company: Anthropic - Website: https://claude.ai/code - Categories: Coding agents with cloud machines - Score: 73/100, #3 of 11 coding agents with cloud machines (method 1.3 (2026-11-14)) - Billing increment: No separate machine billing; sessions use the plan's usage limits - Billed while: The VM is not billed. Model usage in a cloud session counts toward the same five-hour and weekly limits as all other Claude and Claude Code use. - Plans: Pro $20.00/month (Claude Code including cloud sessions. Usage limits every 5 hours and every week, shared across Claude chat, Claude Code and cloud sessions. $17/month when billed annually ($200 up front). Usage credits at API rates after the limit); Max $100.00/month (5x Pro usage every 5 hours and every week, priority access at high traffic times, plus $100 a month in Claude Platform API credits); Ultra $200.00/month (10x Pro weekly usage with no 5-hour limit for verified accounts, plus $200 a month in Claude Platform API credits) - Free tier: None. The Free Claude plan does not include Claude Code; cloud sessions need Pro, Max, Ultra, Team or an eligible Enterprise seat. - Isolation: Isolated Anthropic-managed VM per session - Pause and resume: yes - Snapshots: yes - Custom images: no - Network controls: yes - Max vCPU: 4 - Max memory GiB: 16 - SDKs: Web app, Desktop and mobile apps, CLI - Regions: Not stated. - Open source: no - Self-hostable: yes - Best for: Claude subscribers who want to hand off coding tasks to the cloud at no extra machine cost; Moving work between terminal and cloud with claude --cloud and --teleport; Watching pull requests and auto-fixing CI failures and review comments - Watch out for: Each session is limited to about 4 vCPUs and 16 GB of RAM; Cloud sessions share rate limits with all other Claude usage on the account; Repository cloning and pull requests in hosted environments require GitHub; Paused VMs can be reclaimed, losing running processes - Sources: https://code.claude.com/docs/en/claude-code-on-the-web https://code.claude.com/docs/en/cloud-environments https://claude.com/pricing - Facts checked: 2026-11-14 ## OpenAI Codex cloud Codex Cloud runs coding tasks in OpenAI-hosted VMs that keep working while your computer sleeps, started from the web, mobile, desktop app, CLI or integrations such as Slack and Linear. Each task gets its own workspace built from a reusable environment with repositories, tools, network rules and secrets. The machine has no separate price: cloud tasks draw from the same ChatGPT plan usage allowance as local Codex use, starting with the $20 Plus plan. Plus tasks get 2 vCPU / 8 GiB; Pro, Business and Enterprise get 4 vCPU / 16 GiB. - Page: https://aisandboxhunt.com/openai-codex - Company: OpenAI - Website: https://developers.openai.com/codex - Categories: Coding agents with cloud machines - Score: 69/100, #4 of 11 coding agents with cloud machines (method 1.3 (2026-11-14)) - Billing increment: No separate machine billing; cloud tasks use the plan's usage allowance - Billed while: Not billed by machine time. Usage depends on model, task size, context and tools, and draws from the plan's included limits and then from purchased credits. - Plans: Plus $20.00/month (Codex on the web, CLI, IDE extension and iOS, including cloud tasks. Cloud VM: 2 vCPU, 8 GiB memory, 8 GiB disk. Usage limits reset every five hours and weekly limits may apply; cloud tasks may use more of the allowance than local messages. Extra usage can be bought as credits); Pro 100 $100.00/month (Everything in Plus, with no five-hour limit at present. Cloud VM: 4 vCPU, 16 GiB memory, 32 GiB disk); Pro 200 $200.00/month (Everything in Plus, with no five-hour limit at present. Cloud VM: 4 vCPU, 16 GiB memory, 32 GiB disk); Pro 500 $500.00/month (Everything in Plus plus Ultrafast mode access, with no five-hour limit at present. Cloud VM: 4 vCPU, 16 GiB memory, 32 GiB disk) - Free tier: No free cloud tasks. The Free ($0) and Go ($8/month) ChatGPT plans list Codex only in the desktop app; cloud tasks start with Plus. - Isolation: Each cloud task runs in its own VM - Pause and resume: yes - Snapshots: yes - Network controls: yes - Max vCPU: 4 - Max memory GiB: 16 - SDKs: Web app, Desktop and mobile apps, CLI, Codex SDK - Regions: Not stated. - Open source: no - Self-hostable: no - Best for: ChatGPT subscribers who want cloud coding tasks without paying for machines separately; Teams that need per-environment network allowlists, VPN access or OIDC to cloud accounts; Starting and following tasks from phone, web, Slack or Linear - Watch out for: VM size is fixed by plan: 2 vCPU / 8 GiB on Plus, 4 vCPU / 16 GiB on Pro and team plans; Usage limits are not published as fixed numbers and are shared with all other Codex and ChatGPT Work usage; Not available on the Free or Go plans, nor when using Codex with an API key; Custom base images are not offered; setup is done through install scripts - Sources: https://learn.chatgpt.com/docs/pricing https://learn.chatgpt.com/docs/environments/cloud-environments - Facts checked: 2026-11-14 ## Cursor cloud agents Cursor cloud agents run in their own isolated Ubuntu VM on Firecracker-based microVM infrastructure, with your repositories, dependencies, secrets and a full desktop that the agent can control. Environments are set up by the agent, from a saved snapshot, or from a Dockerfile. The machine has no separate price: cloud agents draw from the plan's usage pool at the selected model's rate, and they need the Premium plan ($30/month) or higher. Idle VMs are hibernated and later deleted; disk snapshots are kept for 30 days of inactivity. - Page: https://aisandboxhunt.com/cursor-cloud-agents - Company: Cursor - Website: https://cursor.com - Categories: Coding agents with cloud machines - Score: 64/100, #5 of 11 coding agents with cloud machines (method 1.3 (2026-11-14)) - Billing increment: No separate machine billing; usage is charged at the selected model's rate - Billed while: Not billed by machine time. Each run draws from the plan's usage pool at the model's token rate; third-party models are billed at API list price. - Plans: Premium $30.00/month (Includes cloud agents and automations. 4x the usage of Plus in a pool shared across Cursor and Grok, refreshed weekly. Third-party models are not covered by the pool and are billed at API list price); Super $100.00/month (15x the usage of Plus, refreshed weekly, plus priority model access at peak times. Third-party models billed separately at API list price); Ultra $200.00/month (50x the usage of Plus, refreshed weekly, plus earliest access to products and models. Third-party models billed separately at API list price); Teams $30.00/month (From $30 per user per month, 4-50x Plus usage depending on seat tier, centralized billing and SSO. Price shown is per user) - Free tier: None. The Free and Plus ($8/month) plans do not include cloud agents. - Isolation: Dedicated VM per agent on Firecracker-based microVM infrastructure, in an AWS account separate from Cursor's other services - Pause and resume: yes - Snapshots: yes - Custom images: yes - Network controls: yes - SDKs: Desktop app, Web and mobile, REST API, Slack, GitHub and Linear integrations - Regions: Not stated. - Open source: no - Self-hostable: yes - Best for: Cursor users who want to hand tasks to the cloud from the editor, web, phone or Slack; Repositories that need a custom Dockerfile-based environment; Agents that must click through a desktop or browser to test their own changes - Watch out for: Not included on the Free or Plus plans; Machine sizes are not published, and only Enterprise can ask for more CPU or memory; Third-party models are billed at API list price on top of the plan fee; Snapshots containing cloned code are kept for 30 days of inactivity and cannot be deleted on demand - Sources: https://cursor.com/pricing https://cursor.com/docs/cloud-agent https://cursor.com/docs/cloud-agent/security https://cursor.com/docs/cloud-agent/setup - Facts checked: 2026-11-14 ## Devin Devin Cloud runs each session on its own machine, booted from a snapshot of an environment you describe in a blueprint, with Linux by default and Windows or macOS VMs as options. Sessions go to sleep after 5 minutes of inactivity and use nothing while asleep. Machine time is not sold separately: it counts toward the plan's usage allowance together with the agent's actions, and Devin says VM time is typically a small fraction of usage. Paid plans get an allowance that refreshes daily and weekly; extra usage is bought as credits consumed at API pricing. Cloud agents start on the $20 Pro plan; the free plan only has a light quota for local agent use. - Page: https://aisandboxhunt.com/devin - Company: Cognition - Website: https://devin.ai - Categories: Coding agents with cloud machines - Score: 64/100, #6 of 11 coding agents with cloud machines (method 1.3 (2026-11-14)) - Billing increment: No separate machine billing; the usage allowance covers agent work and VM time - Billed while: Usage accrues while Devin works; a few units keep the VM running while awake. Asleep sessions use nothing; sessions sleep after 5 minutes of inactivity by default (30 minutes for some Windows or GPU sessions). - Plans: Free $0.00/month (Light quota to code with agents, limited model availability, unlimited inline edits and Tab completions. Access to cloud agents (Devin Cloud) is listed only from Pro up); Pro $20.00/month (Single user. Includes access to cloud agents (Devin Cloud). Usage allowance refreshes daily and weekly and is shared between cloud sessions, Devin CLI and Devin Desktop; up to 10 concurrent sessions. Extra usage past the allowance is bought as credits consumed at API pricing); Max $200.00/month (Single user. Significantly higher quota, refreshed weekly with no daily cap; up to 10 concurrent sessions; extra usage at API pricing); Teams $80.00/month ($80 a month for the team plan plus $40 a month per full seat, as listed on the pricing page. Full seats get a Pro-sized allowance; free flex seats draw from shared extra-usage credits. Up to 100 users, unlimited concurrent sessions, plus 2 concurrent sessions per user in Free mode (SWE-2) at no cost) - Free tier: None for cloud sessions. The $0 Free plan has a light quota for local agent use; access to cloud agents (Devin Cloud) starts with Pro. - Isolation: Each session, including each managed sub-session, runs in its own isolated virtual machine - Pause and resume: yes - Snapshots: yes - Network controls: yes - SDKs: Web app, CLI, REST API, Devin Desktop - Regions: Not stated for self-serve plans. Enterprise can deploy in its own VPC. - Open source: no - Self-hostable: yes - Best for: Handing off whole tickets that need a full machine, browser and desktop to test changes; Teams that want shared environment snapshots built from a declarative blueprint; Projects that need Windows or macOS build machines as well as Linux - Watch out for: No published machine sizes or per-hour machine price; Allowance sizes are not stated in numbers; past the allowance you pay for extra usage at API pricing; Pro and Max allow at most 10 concurrent sessions; The Free plan has no cloud agents; The pricing page and billing docs describe the Teams $80 fee differently - Sources: https://devin.ai/pricing https://docs.devin.ai/admin/billing/self-serve https://docs.devin.ai/admin/billing/usage https://docs.devin.ai/release-notes/2026 https://docs.devin.ai/onboard-devin/environment https://docs.devin.ai/product-guides/security-profiles - Facts checked: 2026-11-14 ## Amp Amp is a coding agent where each new thread can run in its own cloud machine, called an orb, with the repository cloned and tools installed. Orbs come in fixed sizes from 1 vCPU / 2 GB ($0.08/hour) to 16 vCPU / 44 GB, are billed by the minute while running, and pause automatically when idle with files and memory kept. The free Hobby tier pays the standard per-size price for hosted orbs; the $20 Megawatt and $200 Gigawatt tiers include a monthly block of orb time, and self-hosted orbs or your own runners cost nothing on these tiers. Model usage is separate: bring your own subscription or API key, or buy credits at provider list price. - Page: https://aisandboxhunt.com/amp - Company: Amp - Website: https://ampcode.com - Categories: Coding agents with cloud machines - Score: 53/100, #7 of 11 coding agents with cloud machines (method 1.3 (2026-11-14)) - Price per vCPU-hour: $0.0504 - Price per GiB-hour of memory: $0.0162 - Billing increment: per minute - Billed while: Only while the orb is running. An orb pauses 5 minutes after the agent stops working or 20 minutes after your last interaction, whichever is later; a paused orb costs nothing. - Plans: Hobby $0.00/month (No monthly fee and no included orb time: hosted orbs are pay-as-you-go at the standard per-size prices. Self-hosted orbs and your own runners are free and unlimited. Use your own model subscription or API keys, or buy credits); Megawatt (Individual) $20.00/month (750 hours (45,000 minutes) of a1.small (2 vCPU / 4 GB) orbs a month; larger orbs use the hours faster. Also $20 of agent usage for Amp-served models and tools. Allowance resets monthly and does not roll over; after it runs out, buy credits or upgrade. a1.3xlarge is not available); Gigawatt (Individual) $200.00/month (1,000 hours of a1.xxlarge (16 vCPU / 32 GB) orbs a month; larger orbs use the allowance faster and smaller ones more slowly. Also $200 of agent usage for Amp-served models and tools. Allowance resets monthly and does not roll over; after it runs out, buy credits) - Free tier: No free hosted orb time. The Hobby tier itself is free, and self-hosted orbs or your own runners cost nothing on Hobby, Individual and Teams tiers, but Amp-hosted orbs are pay-as-you-go. - Isolation: Separate sandboxed cloud machine per thread; Amp's security page says e2b provides the compute instances. The isolation technology is not stated by Amp. - Pause and resume: yes - Snapshots: yes - Max vCPU: 16 - Max memory GiB: 44 - SDKs: CLI, Web app, Desktop and mobile apps, Plugins (TypeScript) - Regions: Not stated for orbs. Amp's main infrastructure is on Google Cloud in the US. - Open source: no - Self-hostable: yes - Best for: Running many agent tasks in parallel, each in its own paused-when-idle cloud machine; Developers who want to use their own ChatGPT or Claude subscription or API keys for models; Teams that want to run the same agent on their own machines at no extra cost - Watch out for: Hobby has no free hosted orb time; hosted orbs are billed at standard prices; Paid-tier allowances are stated in hours of one orb size, and larger orbs use them faster; The largest size (16 vCPU / 44 GB) is limited to the $200 tier and Enterprise and costs more per resource; The isolation technology of hosted orbs is not documented - Sources: https://ampcode.com/docs/orbs/sizes-and-costs https://ampcode.com/pricing https://ampcode.com/docs/markdown/pricing https://ampcode.com/docs/pricing https://ampcode.com/docs/orbs https://ampcode.com/security - Facts checked: 2026-11-14 ## Factory Factory's Droid agents run locally through the desktop app, CLI and SDK, or in the cloud on Factory-hosted machines: persistent Droid Computers that keep files, packages and running services between sessions, or fresh instances started from a reusable Environment snapshot. Hosted computers auto-pause when idle and resume when a session targets them. There is no separate machine price: compute for cloud environments counts toward the plan's rolling usage limits along with model usage. Plans start at $20 a month; the Free plan is bring-your-own-inference. - Page: https://aisandboxhunt.com/factory - Company: Factory - Website: https://factory.com - Categories: Coding agents with cloud machines - Score: 42/100, #8 of 11 coding agents with cloud machines (method 1.3 (2026-11-14)) - Billing increment: No separate machine billing; compute counts toward the plan's usage limits - Billed while: Not billed by machine time. Sessions consume the plan's standard credits for model usage plus compute used in cloud environments; hosted computers auto-pause when idle. - Plans: Free $0.00/month (Bring your own inference: sign in with ChatGPT, GitHub Copilot or OpenRouter, or connect your own API keys. Whether hosted cloud computers are included is not stated); Pro $20.00/month (Factory-managed inference with rolling 5-hour, 7-day and 30-day usage limits that also cover cloud environment compute. Desktop, CLI and SDK; all products included); Plus $100.00/month (About 5x the usage of Pro); Max $200.00/month (About 10x the usage of Pro); Teams $60.00/month ($60 a month per team plus $40 a month per member, up to 10 seats, each with Pro usage limits and centralized billing) - Free tier: No free hosted compute is stated. The Free plan is free to use with your own model keys or subscriptions, but its access to Factory-hosted cloud computers is not described. - Isolation: Factory-hosted cloud computers; the isolation technology is not stated - Pause and resume: yes - Snapshots: yes - SDKs: Desktop app, CLI, SDK, REST API - Regions: Not stated. Enterprise offers data residency and on-premise deployment. - Open source: no - Self-hostable: yes - Best for: Developers who want persistent cloud computers for an agent that keep state between sessions; Teams that run many parallel agents from a shared environment snapshot; Mixing local, own-server and Factory-hosted machines in one agent workflow - Watch out for: Machine sizes, isolation and limits of hosted computers are not published; Hosted compute shares the same usage limits as model calls, so heavy machine use reduces model headroom; Cloud environment availability depends on your organization's rollout and settings; Usage limits are described only as multiples of Pro, not in fixed numbers - Sources: https://factory.com/pricing https://docs.factory.com/pricing/individuals https://docs.factory.com/droid-computers/overview https://docs.factory.com/environments/overview - Facts checked: 2026-11-14 ## GitHub Copilot coding agent Copilot in the cloud (the Copilot coding agent) runs sessions in an isolated GitHub-hosted Linux sandbox by default, or in a GitHub Actions job when a repository needs custom setup steps, larger or self-hosted runners, or Windows. Unlike most coding agents, the machine is billed separately from AI usage: sandbox compute costs $0.000024 per vCPU-second and memory $0.000003 per GiB-second while running, with snapshot storage for stopped sessions. That compute draws first on the monthly included compute allowance of your GitHub account ($12 on GitHub Free). Cloud sessions need a paid Copilot plan, from $10 a month. - Page: https://aisandboxhunt.com/github-copilot-coding-agent - Company: GitHub - Website: https://github.com/features/copilot - Categories: Coding agents with cloud machines - Score: 42/100, #9 of 11 coding agents with cloud machines (method 1.3 (2026-11-14)) - Price per vCPU-hour: $0.0864 - Price per GiB-hour of memory: $0.0108 - Storage per GiB-month: $0.005 - Billing increment: per second - Billed while: From session start until it is stopped or deleted: compute and allocated memory only while running. Stopped sessions pay snapshot storage until deleted. - Plans: Copilot Pro $10.00/month (Includes Copilot in the cloud and code review, with $15 of GitHub AI Credits a month for model usage ($10 base plus $5 flex). Sandbox compute is billed separately and draws first on the GitHub account's included compute usage: $12 a month on GitHub Free accounts (shown here), $18 on GitHub Pro); Copilot Pro+ $39.00/month ($70 of AI Credits a month, premium models including Opus, delegation to third-party agents. Sandbox compute draws on the GitHub account's $12 (Free) or $18 (Pro) monthly included compute usage); Copilot Max $100.00/month ($200 of AI Credits a month and priority access to new models. Sandbox compute draws on the GitHub account's $12 (Free) or $18 (Pro) monthly included compute usage) - Free tier: None for cloud sessions: Copilot Free does not include Copilot in the cloud. Paid Copilot users get a monthly included compute amount from their GitHub account plan ($12 on GitHub Free, $18 on GitHub Pro) that covers sandbox compute and memory. - Isolation: Isolated GitHub-hosted Linux environment per session; the isolation technology is not stated. Alternatively a GitHub Actions runner job. - Pause and resume: yes - Snapshots: yes - Custom images: no - Network controls: yes - SDKs: Web app, GitHub Mobile, VS Code, Copilot CLI, Copilot SDK, Slack and Microsoft Teams - Regions: Not stated. - Open source: no - Self-hostable: yes - Best for: Teams already on GitHub that want agents working from issues and pull requests; Short tasks where per-second machine billing and the free monthly compute amount keep machine cost low; Repositories that need Windows or self-hosted runners, using the Actions-based option - Watch out for: The sandbox is billed on top of the Copilot plan and AI credits once the account's included compute is used; Sandbox size is not published, and custom compute or setup steps need the Actions option; Actions-based sessions stop after 59 minutes; Not available on Copilot Free - Sources: https://docs.github.com/en/copilot/concepts/security-governance-and-network-settings/about-cloud-sandboxes https://docs.github.com/en/billing/concepts/product-billing/cloud-and-local-sandboxes https://docs.github.com/en/billing/reference/product-usage-included https://docs.github.com/en/copilot/concepts/agents/coding-agent/about-coding-agent https://github.com/features/copilot/plans https://docs.github.com/en/copilot/get-started/plans - Facts checked: 2026-11-14