drydock docs

Operator documentation for drydock: a sandbox for running coding agents (Claude Code, OpenAI Codex, or any OpenAI-compatible model) on your own repos, on macOS, without trusting them. Each task runs in a throwaway VM; the agent never sees your real API key, egress is deny-by-default, and only a diff you approve reaches your code.

New here? Read the Quickstart, install to first task in about a minute.

Pages

• Quickstart: install, then your first sandboxed task.
• Authentication: API key or subscription, for Claude Code and Codex.
• Bring your own model: run any OpenAI-compatible endpoint (Gemini, OpenRouter, local) via opencode.
• Submitting tasks: drydock submit, the approval gate, flags, and scripting.
• Web UI: the board, approval gate, diff, and history in a local browser app.
• Run unattended: brokerd as a launchd agent: starts at login, restarts on crash; spend-limit caveats.
• Egress & widening: the default allowlist, how enforcement works, and per-task widening.
• Configuration: config.yaml reference and env overrides.
• Troubleshooting: drydock doctor and common failures.
• Release QA: the black-box gate run against each installed release, plus the manual browser checklist.
• Threat model: what drydock defends, and what it doesn't.

Requirements

drydock runs on macOS 26+ on Apple silicon: it's built on Apple's container runtime and won't run anywhere else. It is pre-1.0 alpha software with no third-party security audit; the threat model is the contract.