keel / website

# keel

> keel turns coding agents into work owners. It is a project-neutral, multi-agent
> workflow backbone that drives a GitHub issue from intake to done: select →
> branch → implement → CI → review → test → merge → capture → close. Projects set
> their values in `project.yaml` and snap their own logic into named extension
> slots. Stdlib-first Python CLI, single runtime dependency, 100% test coverage,
> 17 `/keel` commands. Installs into Claude Code, Codex, Cursor and Antigravity;
> delegates work to hosted models (Anthropic, OpenAI, Gemini), any OpenAI-compatible
> API, or an agent CLI.

## Docs

- [Home](https://keel-ship.dev/): overview, positioning, and quick start
- [Documentation](https://keel-ship.dev/docs.html): commands, config, extension slots, and the fixed backbone
- [Coverage report](https://keel-ship.dev/coverage.html): live test-coverage dashboard

## Source

- [GitHub repository](https://github.com/berkayturanci/keel): source, issues, and releases

## Key concepts

- Fixed backbone: the invariant step sequence every issue flows through
- Project-neutral core: all project specifics live in `project.yaml`, read via the `keel` CLI
- Extension slots: named hooks where a project snaps in its own implementers/reviewers
- Auditable evidence chain: commit-SHA-bound review provenance, model attribution, and audited deferrals
- Test-first mode: `knobs.implement_mode: tdd` writes the failing tests first, and the blocking `tdd-order` gate verifies that order
- Gate-verified iteration: `knobs.loop` repeats the implementation until the command gates pass, at most `max_iterations` times (a red run after that blocks the issue); under test-first mode it wraps the implementation phase only
- Learning chain: with `learning.enabled`, `learning.mode: create-learning` and a sink, an applied capture writes a Markdown lesson — committed onto the pull request by `keel capture-land` when the sink is inside the repository, written after the merge when it is not — and later implement and review briefs read matching lessons back
- Pure-core / thin-I/O split with a strict 100% coverage bar