Isorun
$19.36 a month*
Home / AI agent sandboxes / Tensorlake Sandboxes
Firecracker microVM sandboxes billed on active CPU time (20% floor), with suspend/resume, snapshots and SSH.
Verdict (2026-11-14): Tensorlake Sandboxes scores 76/100, #5 of 63 AI agent sandboxes. Strong on free tier, isolation, persistence, docs and SDKs. Best for: Agent tool calls and harnesses that wait on model calls, since only active CPU time is billed.
Skip it if: RAM and root disk are billed on allocation for the whole run, and idle CPU still bills 20% of the allocated vCPU.
Prices checked 2026-11-15.
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.
76/100 in AI agent sandboxes, rank 5 of 63. Parts with no data are left out and the rest is rescaled.
Full Tensorlake Sandboxes pricing: plans and costs at four workloads →
$19.36 a month*
$11.73 a month*
$10.14 a month*
$19.00 a month*
$28.44 a month*
$20.00 a month*
$20.00 a month*
$33.44 a month*
| Benchmark | Test | Result | Rank | Data from |
|---|---|---|---|---|
| ComputeSDK benchmarks | Burst start time (median of 100 at once) | 991 ms | 16 of 27 | 2026-11-13 |
| ComputeSDK benchmarks | Repo clone, install and typecheck (one run) | 72,954 ms | 21 of 25 | 2026-11-13 |
| Tensorlake Sandboxes | |
|---|---|
| Category | agent sandbox, dev box, GPU sandbox |
| Our score | 76/100 |
| Isolation | Firecracker microVM per sandbox (GPU sandboxes use gVisor) |
| vCPU per hour | $0.07 (active CPU only) |
| Memory per GiB-hour | $0.015 |
| Storage per GiB-month | $0.07 |
| Billing | per second, billed in whole compute units ($0.01 each), pooled per organization |
| 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. |
| Pause and resume | Yes |
| Snapshots | Yes |
| GPU | Yes |
| Custom images | Yes |
| Network controls | Yes |
| Max vCPU | 16 vCPU |
| Max memory | 64 GiB |
| Open source | No |
| Self-host | Yes |
| SDKs and interfaces | 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. |
| Facts checked | 2026-11-14 |
No reviews yet. Have you used Tensorlake Sandboxes? Your review helps other developers choose.
Tensorlake Sandboxes scores 76/100, #5 of 63 AI agent sandboxes. Strong on free tier, isolation, persistence, docs and SDKs. Best for: Agent tool calls and harnesses that wait on model calls, since only active CPU time is billed. Facts checked 2026-11-14.
$17.95 a month for 2 vCPU, 4 GiB, 176 hours a month (8 h × 22 days), CPU busy 30%. 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.
Firecracker microVM per sandbox (GPU sandboxes use gVisor).
In AI agent sandboxes, the top-rated are Isorun (78/100), Northflank (78/100), Koyeb Sandboxes (76/100). Tensorlake Sandboxes is #5 of 63.
An idle 8-vCPU sandbox bills 1.6 vCPU; a busy one bills what it uses.
Includes 25,000 CU ($250 of usage) per billing cycle. Additional usage is billed at $0.01 per CU at the Pro rates below.
We block new sandbox activity and stop running sandboxes instead of turning the shortfall into an unexpected bill.
The NVIDIA H100 is $3.95 per GPU-hour.
Active CPU per core-hour | $0.07 | $0.042 | Custom
RAM per GB-hour | $0.015 | $0.009 | Custom
Root disk per GB-hour | $0.0002 | $0.0001 | Custom
Snapshot storage per GB-month | $0.07 | $0.07 | Custom
Rates are quoted per hour for readability, but runtime usage is measured by the second.
Pro | $250 / month + usage | 1,000 | 16 vCPU, 64 GB RAM, 100 GB root disk | Unlimited
Free | $0 forever | 1 | 1 vCPU, 1 GB RAM, 10 GB root disk | 2 hours
allocate between 1 GB and 8 GB of RAM per vCPU
GPU sandboxes use gVisor while CPU-only sandboxes run in VMs.
Tensorlake is HIPAA and SOC 2 Type II compliant, supports EU data residency, and offers zero data retention.
Named sandbox is paused. Consumes no compute; state is preserved for resume under the same sandbox ID.
GPU sandboxes are available by request.
Restricting the sandbox's own outbound internet access
Each tool call and harness runs in its own microVM. No shared kernel, no cross-tenant state.