The Pulse
Anthropic Replaces Git With Scion After Merges Top 30,000 a Day
Anthropic says daily changes merged to its monorepo rose from about 2,500 in May 2026 to more than 30,000 in September. The company built Scion, an in-house version-control system, and moved the repository off Git in five months.
AI.info Team · Written with AI ·
Anthropic’s daily code merges jumped from about 2,500 in May 2026 to more than 30,000 in September, prompting the company to move its monorepo off Git and onto an in-house system called Scion. Anthropic says it completed the migration in five months, as AI-assisted coding pushed source control and other developer tools built for human-scale work into the way of its software development.
Why Git became a bottleneck
The scale of Anthropic’s repository helps explain the decision. Its main branch history is almost 100 GB, and a compressed working copy exceeds 25 GB. The company says its continuous-integration systems perform about 25 million code checkouts a day, before counting developer machines, agent sandboxes and research jobs.
Anthropic had tried Git’s partial-checkout features, along with internal mirrors and repository snapshots stored in blob storage. At its peak, it ran several hundred mirrors. Those measures became harder to maintain as the merge rate rose, while snapshots fell further behind the latest main branch and history-dependent operations such as diffs and blame needed extra care.
The company is not arguing that Git is obsolete. “For the vast majority of repositories, this is the right design, and it’s a large part of why git became the standard,” Nick G. and Anthropic’s source control team write. Their point is narrower: a distributed system built to let people work offline can be a poor fit when most clients are always-online agents and jobs accessing one enormous repository.
Scion moves more work to the server
Scion’s clients fetch repository data when they need it instead of holding a complete local copy. The server handles tasks such as creating commits, merging changes and answering questions about file history. Clients keep a bounded cache and can request only selected directories, or materialize files when a job reads them.
That design also lets tools query source control through an API without checking out an entire repository. Anthropic says some jobs can request a diff or perform other work directly through those interfaces. Scion stores changes as changesets that record modified files and point to compressed file contents in blob storage, rather than relying on Git-style packfiles.
Anthropic reports that each tier of Scion’s cache system achieves a hit rate above 99%. It says the new infrastructure uses 200 times less CPU and memory than its former Git-mirroring setup, and that CI and research checkouts now take seconds rather than minutes. The company also reports a 25% CPU-capacity gain for CI alone.
Claude helped build the replacement
Anthropic says a team of seven built Scion, its code-review tool Antique and several related systems while daily merges increased tenfold. The company also describes Claude Mythos Preview rewriting an early Git-to-Scion ingestion pipeline: in Anthropic’s test, the rewritten version used 20 times less memory and ran four times faster than the prototype, while maintaining the ordering and data-structure requirements.
Anthropic’s account does not cast the engineers as spectators. It says they set system goals, made design decisions and redirected Claude when it proposed approaches that violated constraints or added unnecessary complexity. The team also uses internal Claude agents, called “superfans,” to monitor system telemetry and logs alongside conventional alerts.
Scion is not a general Git replacement
Anthropic says Scion is intended for enormous monorepos, not ordinary software projects. Adopting it would also mean migrating code review, search and continuous integration, because Scion diverges significantly from Git and Anthropic’s tools are tailored to its own systems. The company says its unusually high willingness to spend on internal AI use is another reason its design may not suit other businesses.
GitLab responded to Anthropic’s post by describing its own work on source control for agent-heavy development, with a different aim: retaining Git compatibility while changing the architecture beneath it. That contrast points to the unresolved engineering choice for other teams: replace more of the existing toolchain, or adapt Git’s server and client model to handle higher volumes of automated changes.
Anthropic says it wants to share ideas that Git platforms can adopt without requiring every organization to leave Git behind. For now, Scion remains an internal system, and Anthropic’s own account draws a clear boundary around its use: a specialized answer for a repository and workload that the company says most organizations do not yet have.