Claude Container

Run Claude Code autonomously inside a Docker container. The container is the sandbox — full permissions inside, scoped credentials outside.

Quick start

1. Create .env from template: cp .env.template .env (fill in GITHUB_TOKEN, GT_AUTH_TOKEN)
2. Build: docker compose build
3. Run: ./scripts/run.sh (auto-extracts Claude credentials from macOS keychain)
4. One-shot task: ./scripts/run.sh claude -p "Fix the typo in README.md" --max-turns 5

How it works

• Container = sandbox: --dangerously-skip-permissions is safe because the container boundary prevents host damage
• No network firewall: Full outbound network access — WebFetch, npm install, cargo build, pip install, go get all work without restriction
• Repos mounted from host via volume mount: Claude uses --worktree flag to isolate changes — your working tree stays untouched
• Credentials scoped: Fine-grained GitHub PAT + Graphite token injected at runtime
• Safety model: Container isolation + scoped credentials + GitHub branch protection. The container can't touch the host, tokens are least-privilege, and branch protection prevents direct pushes to main

Files

File	Purpose
Dockerfile	Full toolchain: Node, Rust, Go, Python, gh, gt, Claude Code
docker-compose.yml	Volume mounts, env vars, resource limits
entrypoint.sh	Configures git, gh, gt auth from env vars
scripts/run.sh	Extracts Claude credentials from host keychain and launches
claude-config/settings.json	Bypass permissions (container is the sandbox)
claude-config/CLAUDE.md	Autonomous mode instructions
.env.template	Template for secrets

Resource limits

Default: 8GB memory, 4 CPUs. Adjust in docker-compose.yml if cargo/go builds need more.