AI NewsOpen sourceAnnouncement
feder-cr released git-ai, an open-source version of Anthropic's Scion
Federico Elia published git-ai, an MIT-licensed server that implements the ideas in Anthropic's Scion source-control design, built as a fork of GitLab's Gitaly and aimed at teams running hundreds of coding agents at once.
Image: GitHub
Why it mattersA design that was only available to Anthropic engineers can now be self-hosted, so a team running coding agents in parallel can replace the merge queue instead of waiting for a vendor to ship the same thing.
Anthropic described its internal source-control system Scion on 13 October and said it was not available outside the company. Three days later, Federico Elia (feder-cr on GitHub) released git-ai, an MIT-licensed server that reads the same engineering post and implements the ideas it names.
The repository has 1,143 stars at the time of writing and was tagged version 0.1.0, the first public release. The author states on the README that git-ai is "an independent implementation of the publicly described design" and is not affiliated with or endorsed by Anthropic.
What the server does
The README names Scion's starting point: classic Git hosting assumes a human who clones a repository, works for hours and pushes once, while coding agents send thousands of short-lived changes a day, 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 slow that down.
git-ai moves that work to the server, following Scion's design as the author read it. Clients fetch only the objects and paths they ask for. Blame, path history, ancestry and code search run server-side and are memoised as derived data, so the same question is answered once per commit. Landing happens through a compare-and-swap on the branch reference, so concurrent lands are batched into one update instead of serialised through a queue. Every accepted update is a dense, ordered entry in a changeset log that clients can list or watch live. Repository state lives in SQLite or PostgreSQL plus object storage (file, S3, GCS or Azure), and server nodes are caches that any node can rebuild, so a team adds nodes to scale.
A separate graft-verify command tests the branch after a change lands, bisects a failure and reverts the change that broke it. A conflict is handed to a resolver the operator picks: an LLM provider of any kind, a deterministic tool, or none at all. Reviews carry semantic rules, so a comment-only change can need fewer approvals than a logic change, and agents sign in with per-agent tokens that an RPC allowlist and branch policies decide what they may do.
What it costs to adopt
git-ai started as a fork of GitLab's Gitaly, the Go server behind GitLab, and every upstream Gitaly RPC still works. Building it needs Go 1.26, a C toolchain, Meson and Ninja, because the build compiles a bundled Git. The README's own status line says interfaces may still change before 1.0 and names open gaps, including garbage collection of packs in the storage core.
The server never calls a language model itself, so a team can run it against any provider, or against none. Elia asks for issues before large design changes and pull requests with tests, which is the posture of a project meant to be maintained by its community rather than one person. Three days in, the repository has 1,143 stars and the author lists 3,263 followers on GitHub. Whether the Scion ideas work in the open as they do inside Anthropic is a question only running it will answer, and now anyone can.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
