Skip to content

gitai 0.1.0

Latest

Choose a tag to compare

@feder-cr feder-cr released this 15 Oct 13:13
· 7 commits to main since this release

First public release of gitai, an MIT-licensed, vendor-neutral, agent-scale version control server: a Git for coding agents.

gitai is a fork of Gitaly that implements the ideas described in Anthropic's article Redesigning source control. It is an independent project, not affiliated with or endorsed by Anthropic or GitLab. Repositories stay normal Git repositories and upstream Gitaly RPCs keep working.

Highlights

  • Standalone mode for agents: per-agent tokens, an RPC allowlist and branch policies, without GitLab.
  • Landing without a merge queue: UserLand rebases and lands with a compare-and-swap; concurrent lands on one branch are batched; conflicts can be resolved automatically through a vendor-neutral resolver protocol (any LLM provider, or none).
  • Server-side history and data: ObjectService for partial fetches, derived data (graph, changed paths), blame, path history and code search computed on the server.
  • Changeset log: every accepted reference update with its operation and actor, listable and watchable.
  • Review service with semantic approval rules (comment-only changes can need fewer approvals), post-merge verification with automatic bisect and revert, "superfan" monitoring agents, and tiered object caching.
  • Storage core: repository state in SQLite or PostgreSQL as a write-ahead log of changesets with a compare-and-swap per reference, packs in object storage (file, S3, GCS, Azure via gocloud), and server nodes as rebuildable caches. Shadow mode runs it alongside the classic storage; authoritative mode lets any node serve and hydrate any repository.

Getting started

git clone https://github.com/feder-cr/git-ai.git
cd git-ai
make build
examples/graft/quickstart.sh

See the README, the Quickstart and, for the storage core, STORAGE.md.

Known limitations

  • Git still runs on a local copy of each repository; "stateless" means that a node can rebuild that copy from the database and object storage.
  • The storage core has no garbage collection of packs and no checkpoint compaction yet.
  • Authoritative storage mode is not supported behind Praefect.
  • S3, GCS and Azure buckets have not been tested against real services yet (file and in-memory buckets are).
  • Interfaces may still change before 1.0.

The full list is in LIMITATIONS.md.