Skip to main content
Every sandbox has seven coding-agent harnesses, installed and ready: Claude Code, Codex, pi, OpenCode, Prime Agent, Kimi Code, and Mistral Vibe. Run boat prompt to give one of them a task.
1

Connect a key or a subscription

Open Agents & Proxying and sign in to Claude or ChatGPT, or paste an API key. Do this one time. Without it, a prompt fails with 409 provider_not_configured. Read Keys and subscriptions.
2

Give the agent a task

The command prints the work of the agent until the turn ends.
To run the agent on your own GitHub repository, read Run a coding agent on your repo. To run many agents at once, read Run many agents in parallel. boat does these things for you:
  • It streams back what the agent does.
  • It keeps the memory of the agent.
  • It runs many tasks at the same time.
  • It keeps the tasks when the sandbox stops and resumes.
  • It lets you change the model or the harness during a task.
Nothing here is hidden. It is the code you would write if you put an agent on a bare cloud computer. This page shows that code, how boat does it, and where you add your own instructions, tools, keys or harness.

​
What boat builds for you

You can still do each of these by hand. boat never stops you. Run boat ssh and run what you want. Read Bring your own harness.

​
How it fits together

  • Every sandbox runs a small agent server. It is always on.
  • Each conversation has one harness process and one history.
  • The harnesses are ordinary processes. They run as user in /home/user.
  • They read ordinary config files and credentials. So you can look at all of it and replace all of it.

​
Learn more

Harnesses and models

The seven harnesses, their models, the live model catalog, and fast mode.

Keys and subscriptions

Your keys or the keys of your users. Subscriptions, DeepSeek, Amazon Bedrock and Mistral.

AI credentials and proxy

Credential profiles, a second account when one runs out, and placeholder keys for shared sandboxes.

Conversations

Agent memory, parallel conversations, the prompt lifecycle, and a change of harness.

Steer a running turn

Send a new message to an agent while it works, without a stop.

Computer use

The agent sees and uses the sandbox desktop. Tools, config files and checks.

Customize the harness

House rules, MCP servers, skills, tools, and your own harness.

​
Reference