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.
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.
| Scion idea | In gitai |
|---|---|
| Always-online clients that fetch exactly what they need | ObjectService serves single objects and sparse trees; graft clone and graft sparse materialize only the paths you ask for. |
| Source control as an API: history runs on the server | Blame, path history, ancestry and code search run server-side and are memoized as derived data, computed once per commit. |
| Landing without a merge queue | UserLand rebases a change onto the current tip and lands it with a compare-and-swap; concurrent lands on one branch are batched into a single reference update. |
| Automatic conflict resolution | A conflicting land is resolved by a resolver you choose (any LLM provider, a deterministic tool, or none) through a versioned, vendor-neutral protocol. A model's resolution is only a proposal: it lands by itself only if the verify.command you configure passes on the exact tree that would land, and otherwise the agent gets the conflict package with the proposal. The deterministic rules you configure need no such condition of their own, and are verified on the exact tree as well when you turn the verification on (verify.enabled). |
| Changeset log | Every accepted reference update is recorded with its operation and actor in a dense, ordered log that clients list or watch live. |
| Storage core: a database and object storage, stateless servers | Each change is a changeset in a write-ahead log in SQLite or PostgreSQL, with a compare-and-swap per reference; packs live in object storage (file, S3, GCS, Azure). Server nodes are caches that any node can rebuild, so you add nodes to scale. |
| Test after landing, revert the culprit | graft-verify tests the tip, bisects a failure and reverts the change that broke it. |
| Monitoring agents ("superfans") | Agents follow the changeset stream and report what matters to each component, also through Slack. |
| Agent-friendly review | Reviews with semantic approval rules: a comment-only change can need fewer approvals than a logic change. |
| Agents as first-class users | Per-agent tokens, an RPC allowlist and branch policies decide what each agent may do. |
| Strict cache tiers | graft-cache caches immutable objects in daemon and regional tiers, verifying every object. |
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.
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 commandsTo 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.
- 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)
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:
| 50 agents, 250 changes, no conflict resolver | classic | gitai |
|---|---|---|
| changes landed | 100-106 | 200 |
| attempts per landed change | 26.5-28.7 | 1.0 |
| lands per minute | 264-385 | 1,487-1,944 |
| p99 latency | 4.4-5.2 s | 1.9-2.7 s |
The 50 changes gitai does not land are the planned conflicts, returned to their
agents: these runs use no conflict resolver. With the deterministic resolver of
resolve.toml every conflict is resolved
and all 250 changes land (1,245-1,806 lands per minute, p99 2.0-3.4 s). Details and more sizes in BENCH.md.
Version 0.4.0. 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.
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.
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.