Anthropic ditched Git for its own version control system because Claude is writing too much code

Anthropic has moved away from Git for its primary code repository, opting for a custom version control system called Scion. This shift was prompted by the growing volume of AI-generated code changes, which outpaced the capabilities of Git's infrastructure. In May, Anthropic averaged around 2,500 merged code changes daily; by September, this figure soared to over 30,000.

Scion operates most source control operations on servers, significantly reducing the data agents and automated jobs need to download. This architecture cuts CPU and memory usage to just one two-hundredth of what the previous Git mirror infrastructure required. Code checkouts for continuous integration (CI) and research jobs now complete in seconds rather than minutes. The switch has also freed up 25% of the CPU capacity used by Anthropic's CI infrastructure.

Developed by a seven-person team over five months, Scion centralizes commits, merges, rebases, and file-history queries on the server. The client handles local edits, file management, and maintains a limited cache, eliminating the need for each client to store a complete repository. This approach aligns with Anthropic's workload, as most clients are agents and automated jobs that remain online.

Scion's design suits Anthropic's repository size, which approaches 100 GB in its main branch, with compressed working copies exceeding 25 GB. Continuous integration (CI) operations alone perform about 25 million checkouts daily, excluding developer environments, agent sandboxes, and research jobs. Downloading full compressed copies for each CI checkout would generate over 600 petabytes of daily traffic, a scenario Scion avoids.

The company had explored shallow clones, sparse checkouts, promisor remotes, snapshots in blob storage, and internal mirrors before settling on Scion. However, maintaining a network of hundreds of Git mirrors became increasingly challenging as the repository grew. Scion records each change as a changeset in a write-ahead log, with each changeset detailing modified files relative to its first parent and pointing to compressed contents in blob storage.

This design requires resources proportional to the change rather than the entire repository, facilitating higher merge volumes and more effective automated testing.

After merges and before deployment, Anthropic relies more on automated testing, with conflicts typically resolved quickly by Claude, who also contributed to Scion's development. An early internal version of Claude Mythos Preview streamlined a Git-to-Scion ingestion pipeline, running four times faster and using one-twentieth the memory.

While Anthropic acknowledges that Scion may be difficult for other organizations to adopt due to its reliance on custom review, search, and CI tools, they believe existing platforms could benefit from adopting Scion's server-side APIs, particularly those enabling agents to work without repeatedly fetching local copies of the code.