cook

cook coordinates Claude Code, Codex, and OpenCode so developers can give coding tasks repeated reviews or compare competing implementations. It runs as a CLI in a local terminal or CI environment, with a Claude Code skill also available. It requires Node.js and one of the supported coding agents.

Its main distinction is that you can combine workflows. A task can go through several sequential passes, with each pass seeing the previous output, or enter a review loop that checks the work and requests fixes until it meets your criteria. You can assign different agents or models to the implementation and review steps. For longer task lists, an outer loop advances to the next task after the current one passes review.

Parallel runs use isolated git worktrees. You can ask agents to attempt the same task independently or test different approaches, then choose a winning branch, synthesize the results into a fresh implementation, or produce a comparison document without merging code. Each branch can have its own review loop.

The orchestration runs locally; agent requests can still reach external services. cook can use an agent's own OS sandbox or run agents in Docker, where outbound network access is restricted by default to required agent hosts. OpenCode requires Docker mode. Project instructions and per-step agent settings let teams reuse their workflow, and cook automatically waits and retries when an agent encounters a rate limit.