How to run an Agent37 sandbox (via web, CLI, API, or AI agent)
Create an Agent37 sandbox in one command, with no Agent37 account or API key. Or use the dashboard, an AI agent, the API, or the SDK. Median burst TTI 2,360 ms in our Nov 20, 2026 benchmark.

David Tice
Head of Product ·
Agent37 provides persistent Linux sandboxes for AI agents, with signed URLs for any port.
You don't need an Agent37 account, provider SDK, or API key to use it. ComputeSDK runs Agent37 sandboxes on its managed Agent37 account and bills them from your ComputeSDK credit. Sign up once, and you can create an Agent37 sandbox from the dashboard, your terminal, an AI agent, a REST call, or a few lines of TypeScript. If you already have an Agent37 account, you can bring your own key too.
- Dashboard: open Sandboxes → New sandbox and pick Agent37.
- CLI: run
npm create agent37orcompute new agent37. - MCP: connect Claude, ChatGPT, or another agent to
https://platform.computesdk.com/mcp. - REST API: call
POST /api/v1/sandboxesfrom any language. - SDK:
import { compute } from 'computesdk'in a TypeScript app.
How Agent37 performs
ComputeSDK benchmarks Agent37 alongside every other sandbox provider on the same hardware and schedule. Its current results, as of November 20, 2026:
- Sandbox Provider Ratings: ranked #26 of 35
- Burst TTI: 2,360 ms median TTI, ranked #21 of 30
- DAX (dev loop): 131.1 s median dev loop, ranked #25 of 36
- Reliability: 99.999% uptime, ranked #8 of 32
See every result on the Agent37 provider page. Browse the burst TTI, DAX and reliability leaderboards.
Popular comparisons: Agent37 vs Archil, Agent37 vs Arker, Agent37 vs Beam.
Before you start
Create an account at platform.computesdk.com with Google, GitHub, or email. Your account comes with an organization, and every sandbox you create belongs to it.
Sandboxes on ComputeSDK's managed Agent37 account are billed per second from your prepaid credit, which you can add under Settings → Billing. You only pay Agent37 directly if you bring your own key.
1. Create a sandbox in the dashboard
- Sign in to platform.computesdk.com and open Sandboxes in the sidebar.
- Click New sandbox.
- Under Size, pick how much CPU and memory you want.
- Under Routing, pick Agent37.
- Click Create sandbox.

The new sandbox appears in the Sandboxes list, and you can destroy it from there when you're done.
These sizes were available on the managed Agent37 account when we wrote this. The dashboard always shows the current list and the price for each one.
medium: 2 vCPU · 4 GBlarge: 4 vCPU · 8 GB
2. Create a sandbox from the CLI
The fastest way from a terminal is the Agent37 starter:
npm create agent37It offers to install the compute CLI if you don't have it, then runs compute new agent37 for you. The first time, it opens your browser so you can sign in to ComputeSDK and pick your organization. After that, the sandbox is created and ready.

If you already have the CLI, you can do the same in one command:
npm install -g @computesdk/cli
compute new agent37If you aren't signed in yet, compute new opens the browser sign-in first (you can also run compute login yourself). The new sandbox becomes the current sandbox for your directory, so the next commands don't need its ID:
compute exec echo hello # run a command
compute list # see your sandboxes
compute destroy # stop it and stop billing
Run compute new agent37 . to also copy your current project into the sandbox at /app.
3. Let an AI agent create it over MCP
ComputeSDK has a remote MCP server, so an AI agent can create and use sandboxes for you. Connect it to your agent:
- Claude (web and desktop): go to Settings → Connectors → Add custom connector and enter
https://platform.computesdk.com/mcp, or add ComputeSDK Router from the connectors directory. - Claude Code: run
claude mcp add --transport http computesdk https://platform.computesdk.com/mcp. - ChatGPT: install the ComputeSDK Router app from the ChatGPT apps directory.
- Any other MCP client: point it at
https://platform.computesdk.com/mcp.
The first time the agent uses a tool, a ComputeSDK sign-in opens in your browser. Sign in with OAuth, choose the organization the agent should use, and approve the permissions. That's it. Now ask for what you want:
Create an Agent37 sandbox, run
node -vin it, and destroy it when you're done.
The agent creates the sandbox, runs commands, reads and writes files, starts servers, and shares a public preview URL for a port, all through ComputeSDK. Nothing runs on your computer. To disconnect, remove the app under Settings → Connected Apps.
4. Create a sandbox with the REST API
For direct calls from any language, use an organization API key:
- Sign in to the platform and open Settings → API Keys.
- Create a key and copy it. It is shown only once.
- Export it in your shell:
export COMPUTE_API_KEY="your_org_api_key"Create an Agent37 sandbox:
curl https://platform.computesdk.com/api/v1/sandboxes \
-H "Authorization: Bearer $COMPUTE_API_KEY" \
-H 'Content-Type: application/json' \
-d '{"only":["agent37"],"size":"small"}'only limits this create to Agent37. The response's sandbox.id is what you use in the next calls. Run a command, then destroy the sandbox:
curl https://platform.computesdk.com/api/v1/sandboxes/$SANDBOX_ID/commands \
-H "Authorization: Bearer $COMPUTE_API_KEY" \
-H 'Accept: application/json' \
-H 'Content-Type: application/json' \
-d '{"command":"echo hello"}'
curl -X DELETE https://platform.computesdk.com/api/v1/sandboxes/$SANDBOX_ID \
-H "Authorization: Bearer $COMPUTE_API_KEY"
The same API also handles files, background processes, and preview URLs. See the API reference.
5. Create a sandbox with the SDK
In a TypeScript or JavaScript app, install the SDK and use the same API key:
npm install computesdk
export COMPUTE_API_KEY="your_org_api_key"import { compute } from 'computesdk';
const sandbox = await compute.instances.create({
// Follows your routing under Settings → Sandboxes; pick Agent37 there.
size: 'small',
});
try {
const result = await sandbox.runCommand('echo hello');
console.log(result.stdout);
} finally {
await sandbox.destroy();
}Importing computesdk is all it takes: the SDK reads COMPUTE_API_KEY and talks to the platform, so there is no Agent37 package or Agent37 credential in your app. compute.instances.create() follows your organization's saved routing, so pick Agent37 under Settings → Sandboxes to send sandboxes there. The SDK can't pin a single create to ComputeSDK's managed Agent37 account; only the routing setting does that. If you've connected your own Agent37 key, providerOrder: ['agent37'] pins one create to that key, with no fallback.

Bring your own key: try it first, or as a fallback
If you already have an Agent37 account, you can connect its key to ComputeSDK and decide how it is used alongside the managed account. Sandboxes on your own key are billed by Agent37, not from your ComputeSDK credit.
- Add your key under Settings → BYOK.
- Open Settings → Sandboxes and set Use BYOK accounts:
- First: try your own keys first, then ComputeSDK's managed accounts.
- Fallback: use managed accounts first, then fall back to your own keys.
- Only: use only your own keys.
- Off: use only managed accounts.

The setting applies everywhere: the dashboard, CLI, MCP, API, and SDK. You can still override it for a single create, for example compute new agent37 --byok to use your key or compute new agent37 --managed to use the managed account. From the API, set it with PATCH /api/v1/sandboxes/routing and {"byok":"first"}, or pass byok on a single create.
Next steps
Frequently asked questions
Do I need an Agent37 account to run an Agent37 sandbox?
No. ComputeSDK runs Agent37 sandboxes on its managed Agent37 account and bills them per second from your ComputeSDK credit. You can also connect your own Agent37 key.
How do I create an Agent37 sandbox from the terminal?
Run npm create agent37 or compute new agent37. The first time, the CLI opens your browser so you can sign in to ComputeSDK. Then run commands with compute exec and clean up with compute destroy.
Can an AI agent create Agent37 sandboxes?
Yes. Connect Claude, ChatGPT, or another MCP client to https://platform.computesdk.com/mcp and sign in with OAuth. The agent can then create an Agent37 sandbox, run commands, edit files, and destroy it.
Can I use my own Agent37 key with ComputeSDK?
Yes. Add it under Settings → BYOK, then set Use BYOK accounts under Settings → Sandboxes to First to try your key before managed accounts, or Fallback to use managed accounts first and your key as a backup.