---
description: Guide for setting up a sandboxed environment to run Claude Code with recipe-maintainer
allowed-tools: [Read, Glob, Grep]
---

# Setup Sandbox

Explain to the user how to set up a sandboxed environment for running Claude Code with access to recipe-maintainer.

## What to explain

Claude Code needs shell access to run commands, edit files, and install packages. Running it inside a sandbox (like a Docker container) keeps it isolated from your host system while still giving it the tools it needs.

For recipe-maintainer, the sandbox needs:

- **Its own `~/.abra` directory** — so it can manage Co-op Cloud apps and servers without touching your host abra config
- **Access to the recipe-maintainer checkout** — mounted as a volume
- **CLI tools**: `abra`, `git`, `gh`, `ripgrep`, `fd`, and anything else your recipes need (Rust, Hugo, Terraform, etc.)
- **Network access** — to reach Co-op Cloud servers (via Tailscale or direct SSH)
- **Persistent home directory** — so Claude's auth, settings, and conversation history survive across runs

## Reference implementation

Point the user to `sandbox/` in this repo. Read `sandbox/README.md` and summarise the key points:

- `sandbox/` contains a working Docker-based sandbox with all the tools recipe-maintainer uses
- `claude.py` is the launcher — it mounts the current directory, maps your UID/GID, and hides sensitive files
- `docker-compose.yml` defines named volumes for persistence (`claude_userhome`, `claude_target`, `tailscale_state`)
- `entrypoint.sh` creates a non-root user matching the host UID/GID
- The Dockerfile includes all the tools needed for recipe work (abra, Caddy, Tailscale, Rust, Hugo, Terraform, Playwright, etc.)

Tell the user they can either:

1. **Use `sandbox/` directly** — `./build.sh` then set up an alias to `claude.py`
2. **Build their own** — use `sandbox/` as a reference for what tools and configuration are needed

## Key details to mention

- The `.abra` directory is mounted per-workspace via `-v ${workspace}/.container-abra:/home/claude/.abra` so each project can have its own abra config
- Sensitive files (`.env`, `secret.json`, etc.) are automatically hidden from the container via `/dev/null` mounts
- The container runs with `--dangerously-skip-permissions` by default since the sandbox itself is the security boundary
- Auth persists in the `claude_userhome` Docker volume — run `claude login` once and it sticks

Read `sandbox/README.md` for full details and present a clear summary to the user.