Cookwala / For

For developers, robot and appliance makers, and AI-agent builders

Robots and appliances are learning to move. Nobody had written down, in a form a machine can check, what "simmer" means, when chicken is safe, or when a step must be refused. Cookwala is that layer, and it comes with a conformance suite you can run today.

Why it matters to you

Cookwala gives you a task definition and a success test for cooking that is open, royalty-free, model-neutral and device-neutral: recipes a machine can plan, end conditions it can measure, and safety limits it enforces on itself. It sits above motion: your robot learns manipulation its own way; Cookwala tells it what to make, when it is done and what must never happen, and gets back a log.

For agent builders: a mandate model, an untrusted-text rule and a ten-case benchmark that checks whether your agent obeys instructions planted in recipes and listings, plus an MCP server that gives any agent safe, read-only access to recipes and dry runs.

Ways in, light to deep

  1. Light. Run the browser dry run; read Core 0.2 in an evening.
  2. Medium. Install the Python package, dry-run your device's capabilities against the recipes, run the 137 conformance vectors, start the reference hub, add the MCP server to your agent.
  3. Deep. Implement the Core API on a device or hub, publish a conformance report, list your device, write a ROS 2 bridge node, add attack cases to the agent-safety benchmark, propose an RFC.

First success, under 15 minutes

Under 15 minutes, no hardware:

git clone https://github.com/amado2k5/cookwala && cd cookwala && pip install -e sdk/python
cookwala dryrun examples/koshari.cookwala.json --device examples/capabilities/robot-arm.json
cookwala conformance --report report.json
refused: cw.op.boil, missing_capability (no person present) · then 137/137 passed, report.json written

The path after that

  1. Dry run against the example recipes
  2. Implement the Core API against the reference hub
  3. Pass conformance and publish the signed report
  4. List the device in the directory
  5. Consented execution logs become LeRobot datasets and OpenTelemetry traces