Playbooks · Coding tasks
GitAgent: gated commit messages and PRs, locally
Written November 7, 2026
GitAgent drafts commit messages and PR descriptions with a local model and stops at every step for approval. What it gates, and what it costs.
What it does
GitAgent is an agentic graph for commit and deploy workflows. It turns a dirty working tree into an open pull request, one approved step at a time: it scans the repo, drafts a commit message and a PR description with a local ollama model or Mistral, then commits, pushes, and opens the PR. The work is modeled as a graph of steps — scan the tree, draft the commit message, commit, push, draft the PR description, open the PR — and it stops at each one. You see what it proposes and what command it will run before it runs.
The commit message comes from the actual diff, not the branch name: GitAgent scans the working tree and drafts a message describing what changed. The PR description is drafted from the commits on the branch, and the PR is opened through the gh CLI.
What it gates
Nothing touches git history or the remote without an explicit go-ahead on that specific step. The agent proposes; you approve. The gates cover every step that changes state:
- Commit: the exact file list to be committed is shown before anything is staged — untracked files are never swept in.
- Push: the push runs only after approval of that step, with the exact command shown.
- Pull request: the drafted description is shown and the PR opens only after approval.
- Review and merge: a Review → Merge flow can run a second model and focused reviews, and still stops at the merge waiting for approval.
The point of the design is the contrast with the two usual shapes of AI git tooling: a text generator that hands you a commit message to paste, or an agent that runs commands on your repository and tells you afterwards. GitAgent does neither — the model drafts the text, and the steps are code or declared model calls in a flow defined ahead of time, so the model never decides which git commands run.
What "local" means here
Drafting can run entirely on a local ollama model, which means the diff of a private repository never leaves your machine. Mistral is there if you would rather use a hosted model — in which case the diff does leave the machine, and the privacy claim does not apply. The choice is per setup, not per tool: the same install can point at ollama or at a Mistral API key.
GitAgent also runs as an MCP server for Claude Code, Claude Desktop, and Codex, so you can drive it in plain language — for example, asking Claude to run the PR review and merge for an issue. Connecting it that way costs tokens of the agent's own model, not GitAgent AI credits: the server's instructions (about 680 tokens) sit in every conversation, and every reply stays in the conversation and is read again on each later turn.
What it costs
- Free for personal, educational, and non-profit use, in every repository.
- GitAgent Pro is required for commercial use — including a solo consultant on client work and an employee using it at their job. Pro is a monthly plan that also brings AI credits every month for the api0 model. Pro is the commercial license: there is nothing else to buy.
- The source is public and readable under the PolyForm Noncommercial License 1.0.0. The name and logo are trademarks and are not covered by the license.
The current version is 0.1.176, released November 7, 2026. GitAgent ships often; every version and what it changed is listed on the release notes page.
Requirements
- macOS on Apple Silicon, Windows, or Linux x86_64. The macOS build is signed with an Apple Developer ID and notarized.
- git, and the GitHub CLI (
gh) authenticated, for opening pull requests. - Either ollama running locally, or a Mistral API key.
- Pull request creation goes through the GitHub CLI, so PR opening is GitHub-only today. The commit and push steps work on any git remote.
Limits
- This page is a read of the GitAgent product page. No gated commit or PR flow was run for it, and the tool was not installed here.
- The gating and privacy claims above are the product page's description of how the tool works. The approval behaviour is the design; a run would be needed to show what it looks like on a real repository.
- The product page is the only source for the claims above; no independent coverage of the tool was found.
Related playbooks
- Fingerprinting AI coding agents from 33,580 PRs
- Letterpress-1: a growth agent that files PRs
- Add tests to an untested module with an agent
- Addy Osmani's 25 agent skills for coding agents
Sources
Back to: Coding playbooks hub · Also: Manage Git worktrees for parallel AI agents with Worktrunk · Audit a repository's git config before any agent opens it · Run three independent lines of research at the same time in Claude Code, then combine the short summaries