drydock

Let a coding agent run wild on your Mac, without trusting it.

drydock runs Claude Code, OpenAI Codex, or any OpenAI-compatible model (Gemini, OpenRouter, local) full-throttle on your own repos, on your own Mac — no permission prompts, no babysitting. Each task runs sealed in a throwaway VM. The only thing that ever comes back is a git diff, and nothing reaches your real code until you approve it.

• It never gets your key. Your real API key stays on the host; the agent only ever sees a short-lived, budget-scoped token (spend overshoot is bounded to one in-flight request by default).
• It can't smuggle anything out. The internet is deny-by-default: no exfiltrating your code, no calling home (you allow the package registries it needs, nothing else).
• Nothing touches your repo until you say so. You read the diff and approve it before it ever reaches origin.

Most agent tooling tries to keep the agent well-behaved: permission prompts, output filters, policy. drydock takes the opposite stance: contain the blast radius. A hostile agent — a poisoned repo, a malicious dependency, a prompt injection that turns a fetched URL into a shell command — can't reach your key, your filesystem, your push credentials, or the open internet, regardless of what it tries.

_{Don't take the threat model's word for it. Every green above is a real go test red-team case that runs the actual attack and asserts it fails. Reproduce them yourself: see Prove it yourself.}

Docs · Threat model · Website · Roadmap

• Works end-to-end. The full task lifecycle runs (submit → isolated VM → gated diff → push), and drydock ships through a Homebrew tap.
• Containment claims are tested, not asserted. Every claim in the threat model is a red-team test that runs the real attack and asserts it fails. Security-sensitive changes are developed exploit-first and adversarially reviewed.
• Still pre-1.0. Only main is supported, behavior and config can change between minor versions, and it has not been hardened by wide real-world use.
• No third-party security audit yet. The security model is written down in detail in the threat model — read it and decide for yourself before trusting it.
• Hard requirement: macOS 26+ on Apple silicon. drydock runs on Apple's container runtime (1.x, validated through 1.1.0), which ships nowhere else.

• For: unattended or batch agent runs you don't want to babysit, and anyone who'd rather contain an agent than trust it to behave.
• Not for: interactive sessions where you're watching every step (that's overkill), or anything off macOS 26+ Apple silicon (it won't run).

1. Check you're eligible. drydock runs locally, and Apple's container runtime needs macOS 26+ on Apple silicon:

[ "$(uname -m)" = arm64 ] && [ "$(sw_vers -productVersion | cut -d. -f1)" -ge 26 ] \
  && echo "eligible" || echo "not yet, needs macOS 26+ on Apple silicon"

2. Install. Sets up the runtime, squid, and ~/.drydock:

brew install sricola/drydock/drydock
drydock setup

3. Give it a credential and start. A vendor API key works:

export ANTHROPIC_API_KEY=sk-ant-...
drydock start

…or skip keys entirely and reuse a plan you already pay for:

drydock auth claude   # Claude Pro/Max → anthropic_auth: subscription
drydock auth codex    # ChatGPT       → openai_auth: subscription
drydock start

4. Submit your first sandboxed task (in another shell):

drydock submit --repo git@github.com:you/repo \
  --instruction "Add a one-line comment to README.md."
drydock review <id>   # read the diff, then approve or deny

That's the whole loop: the agent runs sealed, hands back a git diff, and nothing reaches your repo until you approve it. The full walkthrough is in the Quickstart.

A few things you'll probably want next:

• Run it unattended (start at login, restart on crash): drydock daemon install — see Run unattended.
• Bound your spend: set aggregate_budget_usd to cap cross-task spend (subscription mode is bounded by task_max_requests instead).
• See run metrics across tasks: drydock stats [--since 30d] [--by agent|vendor|repo|day|week] [--json] aggregates outcome rates, duration and gate-wait percentiles, spend, and egress-widen frequency straight from the audit dir.
• Build from source: brew install go, clone, then make install && drydock init.

Full operator docs live at sricola.github.io/drydock/docs:

• Quickstart: install to first task.
• Authentication: API key or subscription, Claude Code & Codex.
• Models: bring your own OpenAI-compatible endpoint (Gemini, OpenRouter, local).
• Submitting tasks: drydock submit, the approval gate, flags, scripting.
• Web UI: the board, approval gate, and history in a local browser app.
• Egress & widening: the allowlist and per-task widening.
• Configuration: config.yaml and env overrides.
• Security defaults: the shipped financial and containment bounds, generated from code.
• Troubleshooting: drydock doctor and common failures.
• Threat model: what drydock defends, and what it doesn't.

You don't have to trust the threat model — run the attacks. drydock redteam boots throwaway VMs and runs the real red-team cases behind the security claims (key isolation, deny-by-default egress, VM teardown) and prints a pass/fail table. No API key, no spend, ~5 minutes.

drydock redteam    # live containment attacks
make demo VM=1     # …or watch all seven, including live VM isolation

Codebase layout, build/test/CI, and known gaps are in CONTRIBUTING.md. Security reporting and documented residuals are in SECURITY.md.

Apache 2.0, see LICENSE and NOTICE.