Gitmoot

Gitmoot runs a fleet of AI coding agents like a team. It organises agent sessions into seats with roles, scopes and merge authority, keeps them talking through durable messages, directives and wake-ups, and puts their pull requests through an independent review at the exact commit, sized to the risk.

Your agents (Claude Code, Codex, omp, Kimi) write the code in their own sessions. Gitmoot runs on your machine as one static binary with local SQLite state: it watches your repositories, routes reviews to a different agent (on another model if you configure one), and tracks findings until they are answered. Automatic review of every pull request, with a merge gate, is opt-in ([review] native_fanout_enabled). There is no hosted control plane; reviews you choose to send to E2B run there.

:::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.
• Import a cached agent prompt into the current chat with gitmoot agent prompt <agent-or-template>.
• Track jobs, branch locks, review cycles, reviews, and merges locally.
• Run gitmoot dashboard for a styled snapshot of daemon health, repos, agents, jobs by state, worktrees, and branch locks — or gitmoot dashboard --web for a read-only browser view of a running orchestration.

How It Works

Codex, Claude Code, Kimi Code, and omp 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