# Use from your coding agent

Your coding agent on your laptop, such as Claude Code, Codex or Cursor, can drive HareBox through the CLI: start machines, run builds and tests in them, fork them, and stop them, while your laptop stays quiet.

## 1. Give it the docs

Add HareBox's docs as an MCP server, so the agent searches and reads current docs instead of guessing:

```sh
claude mcp add harebox-docs -- harebox mcp docs
```

For other agents, register `harebox mcp docs` as a stdio MCP server. It has two tools, `search_docs` and `read_doc`, needs no key, and reads the published docs, so it's always current.

## 2. Teach it the rules

- Copy [`AGENTS.template.md`](../agents/AGENTS.template.md) into your repository as `AGENTS.md`, or into `CLAUDE.md`.
- Install the [HareBox skill](../skills/harebox/SKILL.md) where your agent loads skills.
- Point it at [`llms.txt`](../llms.txt). Everything it needs is reachable from there.

## 3. Let it work

The CLI prints JSON whenever its output is piped, and every error has a `fix` that says what to do next, so an agent can run it unattended:

```sh
harebox help --json | jq '.commands | length'
```

Typical requests that work well:

- "Start a large HareBox sandbox, push this repository to it, run the full test suite there and tell me what fails."
- "Fork the sandbox three times and try three fixes for the flaky test in parallel; keep the one that passes."
- "Run `harebox run --from ci-base -- make test` and report the exit code."

Set a budget first, so a loop can't run up a bill: `harebox budget new --limit 25 --action stop`.