Git designed for AI agents.

Anthropic built Scion, a faster version control system for coding agents, and described it in Redesigning source control: a server-centric system built for hundreds of agents that read, write and land code at the same time. Scion is internal to Anthropic and is not available to the public.

So I built gitai: an open-source version of that design that anyone can run, and that the community can improve and maintain together.

MIT licensed. Vendor neutral: the server never calls a language model, and every AI feature works with any provider, or none.

_{gitai is an independent implementation of the publicly described design. It is not affiliated with or endorsed by Anthropic.}

Why

Classic Git hosting was built for humans who clone a repository, work for hours and push once. Coding agents work differently: thousands of short-lived tasks, each touching a few files, each wanting to land in seconds, all on the same branches. Full clones, merge queues and client-side history walks become the bottleneck.

Scion's answer is to move the work to the server: clients stay online and fetch only what they need, history questions are answered (and cached) by the server, changes land through a compare-and-swap instead of a queue, every update is an entry in a changeset log, and the repository state lives in a consistent database plus object storage so servers scale out. gitai implements these ideas, with normal Git repositories underneath.

The Scion ideas, implemented

Architecture

        agents, CI jobs, sparse workspaces          graft CLI / Go SDK
                            |
                            v   gRPC
        +-------------------------------------------------+
        |  graft-cache (optional daemon and cache tiers)  |
        +------------------------+------------------------+
                                 v
   +---------------------------------------------------------------+
   |  gitai server (any number of nodes)                           |
   |  land - objects - changesets - derived data - search - review |
   |  write path: changeset + compare-and-swap per reference       |
   +-----------+--------------------------------+------------------+
               |                                |
               v                                v
   +-----------------------+       +-----------------------------+
   |  SQLite / PostgreSQL  |       |  object storage             |
   |  changeset WAL, refs, |       |  content-addressed packs    |
   |  changelog, reviews   |       |  (file, S3, GCS, Azure)     |
   +-----------------------+       +-----------------------------+
               \                              /
                local Git repositories = caches any node can rebuild

Details: ARCHITECTURE.md and STORAGE.md.

Quick start

Requirements: Linux or macOS, Go 1.26, a C toolchain, Meson and Ninja (the build compiles a bundled Git).

git clone https://github.com/feder-cr/git-ai.git
cd git-ai
make build                         # server, CLI and tools into _build/bin
examples/graft/quickstart.sh       # a server, a cache daemon, two agents and a land, in a temporary directory
./_build/bin/graft                 # the CLI and its commands

To run your own server, adapt examples/graft/config.toml and start it with _build/bin/gitaly serve config.toml. To put the repository state in a database and object storage, add a [graft.storage] section: start in shadow mode, import your repositories, then switch to authoritative (STORAGE.md).

Inside the code the extensions are called Graft: the [graft] configuration section, the graft CLI, internal/graft and doc/graft.

Documentation

• Quickstart and configuration reference
• Architecture and the storage core
• Landing and batching, conflict resolution, derived data
• Review, search, verification, superfans and Slack chat
• Agent security, benchmarks, fuzzing
• Limitations and roadmap
• Design documents of each phase: PLAN.md to PLAN6.md, and PLAN8.md (faster landing)

Benchmark

50 agents land 5 changes each on one branch; 20% of the changes conflict on purpose. "Classic" is the same server with batching off: every agent rebases and retries its own compare-and-swap, up to 20 attempts. Measured with acceptance-phase4.sh on one 16-core machine, two runs:

The 50 changes gitai does not land are the planned conflicts: these runs use no conflict resolver. With the deterministic resolver of resolve.toml every conflict is resolved and all 250 changes land. Details and more sizes in BENCH.md.

Status

Version 0.2.0. Everything above is implemented and covered by unit, end-to-end, multi-node, crash and fuzz tests. Interfaces may still change before 1.0. Known gaps (for example garbage collection of packs in the storage core) are listed in LIMITATIONS.md.

Community

gitai is meant to be maintained by its community. Open an issue to report a bug or propose an idea, start a discussion before a large design change, and send pull requests: tests and a short description of the reasoning make them easy to review. The design documents in doc/graft/PLAN*.md explain why things are the way they are.

Credits and license

gitai is MIT licensed, see LICENSE.

It started as a fork of Gitaly, GitLab's Git RPC server (MIT, Copyright GitLab Inc.), which provides the Git plumbing underneath; every upstream Gitaly RPC keeps working. The non-MIT ee/ directory of upstream Gitaly is not part of this repository. Gitaly's own README is kept in doc/UPSTREAM-GITALY.md, and third-party notices are in NOTICE.