cook

Do you often find yourself in this loop?

You:   Hey Agent, implement dark mode.
Agent: Done! I added the thing.

You:   Hey Agent, review your work.
Agent: Found a few issues.

You:   Hey Agent, fix your work.
Agent: Fixed! 

Issue the task once and let it cook:

cook "Implement dark mode"

cook runs Claude, Codex, or OpenCode in a sandboxed work → review → gate loop, iterating automatically until the agent is satisfied or your max iterations are hit. Get even fancier by defining what to review and the criteria for done:

cook "Implement dark mode" \
  --review "Review the implementation. Check for visual regressions and missing theme variables. Categorize findings by High/Medium/Low." \
  --gate "Reply DONE if no High findings remain; otherwise ITERATE." \
  --max-iterations 5

Prerequisites

• Node.js 20+
• Docker

Install

npm install -g @rjcorwin/cook

Quick start

cd your-project
cook init                      # creates COOK.md, .cook.config.json, .cook.Dockerfile
cook doctor                    # checks Docker + auth readiness
cook "Add dark mode"           # runs the agent loop
cook "Fix the login bug" 5     # runs with up to 5 iterations

Examples

Basic run:

cook "Create a space cat themed todo app in a single index.html"

Per-step agent/model overrides (for one run):

cook "Implement auth flow" \
  --agent opencode --model gpt-5 \
  --work-agent codex --work-model gpt-5-codex \
  --review-agent claude --review-model opus

Doctor checks using the same override logic:

cook doctor --work-agent codex --review-agent claude

How it works

1. Work — The selected agent executes your prompt inside a Docker sandbox with the project bind-mounted.
2. Review — A second pass reviews what changed and flags issues by severity.
3. Gate — A third pass decides DONE or ITERATE based on the review.

The loop repeats until the gate says DONE or max iterations are reached (default: 3).

A persistent status bar at the bottom of the terminal shows the current step, iteration, model, and elapsed time.

Sandbox security

The agent runs inside a Docker container — it can freely read and write your project files, but it cannot touch anything else on your host machine.

Network access is restricted by default using iptables inside the container. Only outbound HTTPS to the agent's API endpoint is allowed (e.g. api.anthropic.com for Claude). Everything else — including Google, npm, GitHub, etc. — is blocked unless explicitly added to allowedHosts.

To allow additional hosts:

{
  "network": {
    "mode": "restricted",
    "allowedHosts": ["registry.npmjs.org", "api.github.com"]
  }
}

To disable restrictions entirely (not recommended):

{
  "network": {
    "mode": "unrestricted"
  }
}

Configuration

• COOK.md — Project instructions and agent loop template (JS template literal syntax).
• .cook.config.json — Default agent/model, per-step agent/model overrides, network restrictions, and environment variable passthrough.
• .cook.Dockerfile — Project-specific dependencies layered on top of the base sandbox image.

Example .cook.config.json:

{
  "agent": "opencode",
  "model": "gpt-5",
  "steps": {
    "work": {
      "agent": "codex",
      "model": "gpt-5-codex"
    },
    "review": {
      "agent": "claude",
      "model": "opus"
    },
    "gate": {}
  },
  "network": {
    "mode": "restricted",
    "allowedHosts": []
  },
  "env": ["CLAUDE_CODE_OAUTH_TOKEN"]
}

CLI defaults (--agent, --model) override config defaults for a single run. Step flags (--work-agent, --review-agent, --gate-agent, --work-model, --review-model, --gate-model) override both.

Options

cook "prompt"                   Run the work/review/gate loop
cook "prompt" 5                 Run with 5 max iterations
cook init                       Set up COOK.md, config, and Dockerfile
cook rebuild                    Rebuild the sandbox Docker image
cook doctor                     Check Docker + auth readiness

--work PROMPT                   Override work step prompt
--review PROMPT                 Override review step prompt
--gate PROMPT                   Override gate step prompt
--max-iterations N              Max review iterations (default: 3)
--agent AGENT                   Default agent (claude|codex|opencode)
--model MODEL                   Default model (for default agent)
--work-agent AGENT              Work step agent override
--review-agent AGENT            Review step agent override
--gate-agent AGENT              Gate step agent override
--work-model MODEL              Work step model override
--review-model MODEL            Review step model override
--gate-model MODEL              Gate step model override