Skip to main content

Gitmoot

Gitmoot coordinates local AI agents through the same surface teams already use to audit software work: repositories and pull requests.

It runs on the user's machine, stores workflow state in local SQLite, polls GitHub pull request comments, routes jobs to registered agent runtimes, and writes attributed results back to the pull request discussion. There is no hosted control plane in the current beta.

:::tip For agents New here? Open llms.txt — the machine-readable index of these docs — and the agent skill SKILL.md. llms-full.txt has the full expanded context in one file. :::

What Gitmoot Is For​

  • Route PR comments to named local agents.
  • Keep Codex, Claude Code, Kimi Code, shell, and future runtimes behind one agent model.
  • Start or subscribe agents with explicit repo access and capabilities.
  • Orchestrate work across multiple agents with structured delegations and automatic synthesis.
  • Use agent templates for reusable planner, review, or custom prompt agents.
  • Capture a successful current chat as a reviewed, reusable agent template draft.
  • Import a cached agent prompt into the current chat with gitmoot agent prompt <agent-or-template>.
  • Track jobs, branch locks, goals, tasks, reviews, and merges locally.
  • Open gitmoot dashboard, an interactive TUI cockpit with pages for Attention, Activity (live orchestras), Trains, Agents, Workers, Jobs, Locks, Health, and Config (pending prompts live under Attention) — or gitmoot dashboard --web for a read-only browser view of a running orchestration.
  • Run SkillOpt training loops that collect human feedback, call the external optimizer, and keep promotion decisions explicit.

How It Works​

Codex, Claude Code, and Kimi Code are the runtimes Gitmoot can start or subscribe; shell is subscribe-only.

Use the docs here for the human workflow; agents have the index callout above.

Docs Map​