Sanbox By Sanlabs

Guides

Building agent products

Available on request

Use Sanbox from any MCP client

By the Sanbox team updated

Ask ChatGPT Ask Claude

The Sanbox MCP server lets any MCP client start Sanbox Computers, run commands in them, move files and collect results, all as tools the client can call. It's enabled per organisation on request. Today, any agent with shell access does the same through the Sanbox CLI, and sanbox skill gives it the full instructions.

What you get

The Model Context Protocol (MCP) is an open standard for connecting AI applications to external tools and data. Assistants such as Claude and ChatGPT, and development tools such as Visual Studio Code and Cursor, support it.

  • Computers from any MCP client. A chat assistant without a terminal can start a Computer from one of your templates, work in it and bring back the results.
  • The same Computers. Every Computer still comes from your templates, so its network policy, Secret Proxies and disk work exactly as they do when you start it from the CLI.
  • No terminal needed. People who never open a shell can use Computers from the MCP client they already work in.

Enable it for your organisation

Write to sales@sanlabs.ai with the subject "Enable: Sanbox MCP server", and tell us which MCP clients you use.

What works today: the CLI

Coding agents with a shell, such as Claude Code, Codex and OpenCode, already drive Sanbox through its CLI. Install it, sign in, and save the agent guide into your project:

npm install -g @sanlabs/sanbox-cli@latest
sanbox login
sanbox skill --write
sanbox templates list --json

sanbox skill --write saves the current guide to .sanbox/agent.md. It covers templates, runs, Computers, files, ports, Secret Proxies and recovery. Point your agent's instructions file at it, or just ask:

Read .sanbox/agent.md. Start a Computer from the "Compose stack" template,
copy ./app into it, run the end-to-end tests there and bring back the report.
Show me each command before you create, stop or delete a Computer.

The CLI is built for agents:

  • JSON everywhere. --json returns one envelope with ok, data or error, and next_actions the agent can follow as argument arrays.
  • Safe retries. --external-run-id on runs and --request-key on forks, commands and snapshots stop a retry from doing the work twice.
  • Clear exit codes. 0 for success, 1 for a local or API failure, 2 when readiness fails or a waited run ends unsuccessfully.

For automation, set an organisation API key in SANBOX_API_KEY instead of signing in. Keep it out of prompts, task files and repositories.

Can a Computer use my own MCP servers?

That's the other direction, and it works today: you bring your own models, tools and MCP servers to the agent in a Computer. Allow the MCP server's host in the template's network policy. If the server is reached over HTTPS with a token, a Secret Proxy can hold the token so the agent never sees it.

Sources