Oak: a content-addressed version control system built for AI coding agents

Oak is a public-beta (v0.99.0) version control system pitched as a Git alternative optimized for AI coding agents rather than humans. Instead of human conventions like per-commit messages, it organizes work around a branch-per-session model with branch-level descriptions, and uses content-addressed storage with lazy, on-demand mounting so an agent can start editing any repository within seconds. The project frames its speed advantage over Git as a side effect of this design rather than its main selling point, and it’s meant to slot underneath existing tools like Claude Code, Codex, or Cursor as the layer they read, write, branch, and collaborate through.

The codebase is a Rust Cargo workspace split into a reusable library (published as oakvcs-core, imported as oak_core) and the oak command-line client that agents drive. The core crate can be pulled in standalone for just the content-addressed data model and hashing, with an optional local-repo feature adding on-disk SQLite and read-only Git backends. Distribution covers macOS (Apple Silicon), Linux (x86_64), and Windows, where the mount feature relies on the Projected File System (ProjFS); installation is available via a curl script, prebuilt binaries, or cargo install. It’s Apache-2.0 licensed.

Notably, the maintainers state the repository was written almost entirely by AI with human oversight — a fitting origin for a tool whose entire premise is that version control should be shaped around how agents actually operate. Whether agent-native VCS becomes a real category or stays a niche optimization, Oak is an early concrete bet that the assumptions baked into Git don’t map cleanly onto autonomous, session-based machine workflows.