Groundbox vs Tensorlake for agent sandboxes.
Both run sandboxes for agents as Firecracker microVMs that keep memory when paused, and both also run GitHub Actions jobs. Pick Tensorlake for memory snapshots that start many warm copies, CPU billed only while busy, GPU sandboxes, or a SOC 2 Type 2 report. Pick Groundbox for memory included in the CPU price, a lower price per hour while the CPU is busy, and a spend cap and scopes on every API key.
Tensorlake bills CPU only while busy; the bar is the busy end.
sources: groundbox.dev/pricing · docs.tensorlake.ai · checked 2026-11-20
Side by side
| Feature | Groundbox | Tensorlake |
|---|---|---|
| Price, 2 vCPU · 4 GB | $0.08 / hour | $0.088 to $0.20 / hour |
| Pause keeps memory | Yes, memory and disk | Yes, memory and running processes |
| CI runners too | Yes, on the same key and bill | Yes, GitHub Actions runners |
| Isolation | Firecracker VM | Firecracker microVM |
| Fork and snapshots | Fork a running sandbox; save points | Memory snapshots and warm copies |
| Longest session | 720 hours, renewable | No limit on Pro; 24 hours on credits, 2 hours free |
| Self-host | No | In your VPC or on-prem, on Enterprise |
| GPUs | Not offered | H100, by request |
| Largest size | 16 vCPU | 16 vCPU on Pro |
| SDK languages | TypeScript and Python | Python and TypeScript |
| Security report | No SOC 2 report | SOC 2 Type 2; HIPAA on Pro |
| Free to start | $5 of usage, no card | 1 sandbox, 2-hour sessions, no card |
Tensorlake runs sandboxes for agents as Firecracker microVMs with memory snapshots, and also runs GitHub Actions jobs. A rust mark means that side is ahead. Where they are ahead, we say so. Every value of theirs links to the page it is from; “Not published” means their pages do not say. More comparisons: all comparisons.
Who each is for
Pick Tensorlake for
- memory snapshots that start many warm copies of one sandbox
- CPU billed only while busy, with a floor of 20% of the cores
- GPU sandboxes (H100, by request), and in-VPC or on-prem on Enterprise
- a SOC 2 Type 2 report on every tier, and HIPAA on Pro
Pick Groundbox for
- memory included in the CPU price ($0.04 per vCPU-hour), on dedicated cores
- pause, fork and save points of a running sandbox, and an egress allowlist
- a spend cap and scopes on every API key
Every detail
| Feature | Groundbox | Tensorlake |
|---|---|---|
| Plans | Free: everything, with $5 of usage. $20 / month: $20 of usage included, then pay as you go | Free; prepaid usage credits from $5; Pro $250 / month |
| CPU | $0.04 per vCPU-hour, a dedicated core | $0.07 per active core-hour on credits, $0.042 on Pro; at least 20% of the cores is billed |
| Memory | Included, up to 4 GB per vCPU | $0.015 per GB-hour on credits, $0.009 on Pro, billed apart from CPU |
| Storage | 20 GB disk by default, up to 100 GB | Root disk $0.0002 per GB-hour; snapshots $0.07 per GB-month |
| Billing | Per second, no minimum | Per second |
| Sizes | 1 to 16 vCPU, up to 4 GB per vCPU | Up to 4 vCPU on credits and 16 on Pro, 1 to 8 GB per vCPU |
| Default size | 2 vCPU, 4 GB RAM, 20 GB disk | 1 vCPU, 1 GB RAM, 10 GiB disk |
| Longest session | 720 hours, renewable; 60 minutes by default | 2 hours free, 24 hours on credits, no limit on Pro |
| Sandboxes at once | 20 once a card is on file | 1 free, 100 on credits, 1,000 on Pro |
| Isolation | A Firecracker VM, a new one per job or sandbox | Firecracker microVMs; GPU sandboxes use gVisor |
| Pause and resume | Yes, memory and disk; $0.05 per GB-month while paused | Named sandboxes keep files, memory and running processes |
| Fork and snapshots | Fork a running sandbox with its files and processes; save points to rewind to | Filesystem or memory snapshots; many warm copies of one sandbox |
| Egress controls | An allowlist of hosts per sandbox | Allow and deny lists by domain, IP or CIDR, changeable while running |
| GPUs | Not offered | H100, by request |
| SDKs | TypeScript and Python, plus a REST API, CLI and MCP server | Python and TypeScript, plus a CLI |
| CI runners | GitHub Actions runners on the same key and bill | GitHub Actions runners, picked with runs-on: tensorlake-medium |
| Self-host | No | In your VPC or on-prem, on Enterprise |
| Security reports | No SOC 2 report | SOC 2 Type 2 on every tier; HIPAA with a BAA on Pro |
A month at list price
One sandbox with 2 vCPU and 4 GiB of memory, running 100 hours in a month:
| Groundbox | $8.00 ($0.04 per vCPU-hour, memory included). On the $20 plan it fits in the $20 of usage included. |
|---|---|
| Tensorlake | $8.80 to $20.00, from CPU at its 20% floor the whole time to CPU busy the whole time |
At Tensorlake's usage-credit rates. Its Pro rates are lower ($0.042 per core-hour, $0.009 per GB-hour), with $250 a month on top. Disk is billed apart and not counted. All our prices are on the pricing page.
Speed
We have not measured this service's sandboxes, and we publish no start or fork times for ours, so this page makes no speed claim. The speed figures we do publish are for CI runners, on the benchmark page. Method, scripts and every project's times: the benchmark page.
Switch in four steps
- Create an API key in the dashboard. An agent can also sign itself up with groundbox signup and send its person a claim link.
- Install an SDK: npm i @groundbox/sdk or pip install groundbox.
- Where your code creates a sandbox and runs a command in it, call create_sandbox and run_command. Files, fork, pause and stop have one method each, with the same names as the MCP tools.
- Or give the agent the MCP server instead of SDK code: it has the same tools.
Details in the docs: First sandbox, the SDKs and the MCP server.
import { Groundbox } from "@groundbox/sdk";
const groundbox = new Groundbox(); // reads GROUNDBOX_API_KEY
const sandbox = await groundbox.create_sandbox({ cpus: 2 });
const result = await groundbox.run_command(sandbox.id, "uname -a");Questions
| Is Groundbox cheaper than Tensorlake? | At list price, one 2 vCPU, 4 GiB sandbox for 100 hours costs $8.00 on Groundbox and $8.80 to $20.00 on Tensorlake, before plan fees and credits. Tensorlake bills CPU only while busy, so a mostly idle sandbox can cost about the same as on Groundbox. |
| How do I move an agent from Tensorlake to Groundbox? | Create an API key, install the TypeScript or Python SDK, and call create_sandbox and run_command where your code makes a sandbox and runs a command. The MCP server has the same tools. |
| How are sandboxes isolated on Groundbox and on Tensorlake? | Each Groundbox sandbox is its own Firecracker VM on dedicated cores. Tensorlake CPU sandboxes are Firecracker microVMs too; its GPU sandboxes use gVisor. |
| Can a Groundbox sandbox pause and fork? | Yes. Pause keeps memory and disk and costs $0.05 per GB-month. Fork copies a running sandbox's files and processes. |
| How long can a Groundbox sandbox run? | It stops after 60 minutes by default. The longest timeout is 720 hours, and it can be renewed. A Tensorlake session lasts up to 2 hours on the free tier and 24 hours on usage credits, with no limit on Pro. |
| Does Groundbox also run CI? | Yes. GitHub Actions runners from $0.003 a minute are on the same API key and bill as sandboxes. |
Sources
Tensorlake's facts come from its own pages, read on the day shown. Prices change; their pages are the final word.
- Tensorlake pricing, checked 2026-11-20
- Tensorlake docs, checked 2026-11-20
- Tensorlake docs: billing, checked 2026-11-20
- Tensorlake docs: sandbox lifecycle, checked 2026-11-20
- Tensorlake docs: snapshots, checked 2026-11-20
- Tensorlake docs: networking, checked 2026-11-20
- Tensorlake docs: GPU sandboxes, checked 2026-11-20
- Tensorlake docs: quickstart, checked 2026-11-20
- Tensorlake docs: GitHub Actions, checked 2026-11-20