gitai CLI
gitai is an open source-control server for hundreds of agents: landing by compare-and-swap, server-side history and the graft CLI.
- 1,013 ★ on the repository
- Last commit
Extended access
Why this level
- The server accepts changes from several agents into shared branches and can revert commits itself.
- Agent rights are set through tokens and policies, and a misconfiguration gives an agent excess access.
Install
- Works in
Manual install
make buildAfter `git clone https://github.com/feder-cr/git-ai.git`. Then `examples/graft/quickstart.sh`. Go 1.26, a C compiler, Meson and Ninja are required.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The server accepts changes from several agents into shared branches and can revert commits itself.
- Agent rights are set through tokens and policies, and a misconfiguration gives an agent excess access.
This is third-party code. Review the repository files before installing.
What it does
gitai is an independent open implementation of the ideas of Scion, a source-control system for coding agents described by Anthropic in the article Redesigning source control. The author stresses the project is not affiliated with Anthropic. Underneath are ordinary Git repositories, with work moved to the server: clients stay online and fetch only needed objects, and history, blame and code search are computed on the server and cached. Changes land without a merge queue, through compare-and-swap on a reference. A conflict is resolved by a tool you choose: any model, a deterministic utility or none. Every update is written to a changeset log. `graft-verify` tests the branch tip, bisects to the breaking change and reverts it. Each agent has its own token, an allowlist of calls and branch policies. The CLI is called `graft` and the server starts with `_build/bin/gitaly serve config.toml`. In the author benchmark with 50 agents and 250 changes, 200 land against 100 to 106 in classic mode.
Who it is for. Infrastructure and platform teams where dozens of agents land changes on the same branches at once.
Good fit when
- Fifty agents edit one branch and the usual merge queue becomes a bottleneck.
- An agent needs a sparse clone: only the paths it works on, not the whole repository.
- A conflict between two agents must be resolved by a model or a deterministic script.
- After landing, the branch tip must be tested and the breaking change reverted automatically.
Not a fit when
- One person and one agent work: ordinary Git and pull requests are enough.
- You lack Linux or macOS, Go, Meson and Ninja: the build compiles a bundled Git.
- You need a hosting service with a web interface and human accounts: gitai does not replace GitHub.
Example request
Run the gitai quickstart and show how two agents land changes on one branch and what happens on a conflict.Limitations
Building needs Linux or macOS, Go 1.26, a C compiler, Meson and Ninja because the build compiles a bundled Git. It is an independent implementation of a publicly described idea and not an Anthropic product. README benchmarks were run by the author on one 16-core machine and repeated twice for version 0.3.0, and not repeated for 0.4.0. Storing state in a database and object storage is enabled in steps: `shadow` mode first, then `authoritative`. The GitHub API reports no detected license while the project text names MIT. The last commit dates from November 2026.
How to disable. Stop the `gitaly` server and the cache daemon, then delete the `_build` directory. Data lives in ordinary Git repositories and the chosen storage, so keep or delete it separately.
FAQ
How do I build gitai?
Clone the repository and run `make build`: the server, CLI and tools appear in `_build/bin`. Go 1.26, Meson and Ninja are required.
How do I see it working quickly?
Run `examples/graft/quickstart.sh`: it starts a server, a cache daemon and two agents in a temporary directory and performs a land.
Who resolves conflicts?
A resolver you choose: any model, a deterministic program or none. The rules are in `doc/graft/RESOLVE.md`.
Is the project affiliated with Anthropic?
No. gitai is an independent open implementation of the publicly described Scion design, and the author notes no affiliation with Anthropic or OpenAI.
What protects agents from each other?
Per-agent tokens, an allowlist of calls and branch policies. Details are in `doc/graft/AGENT-SECURITY.md`.
Discussion
No questions or reviews yet.
Related
CLI toolsA skill that splits changes into logical Conventional Commits and safely handles branches, pushes and pull requests.
multi-gitter runs one script across multiple Git repositories and creates pull requests for the resulting file changes.
A CLI and skill for precise non-interactive git: commit single hunks, split commits and fold fixes into earlier ones.
A Git merge driver and MCP server that compares functions and classes and removes false conflicts from parallel agent edits
A command-line SSH tool for agents: remote execution, file upload and download, and command allow and block lists.
A CLI that installs shared skills, commands, subagents, hooks and MCP servers into eight AI coding tools in their native layouts
README in short
A Go project around Git with a server part, the `graft` CLI, a cache daemon and verification tools. Docs on storage, review, search and agent security are in `doc/graft`.