Anthropic replaces Git with custom source control system Scion

Anthropic has migrated its main code repository from Git to an internally built version control system after Claude helped drive daily code merges from roughly 2,500 in May to more than 30,000 in September.

The company said a seven person team built the new system, called Scion, alongside a code review tool and other developer infrastructure, completing the migration in five months.

The project reflects a shift in where Anthropic encounters limits as AI agents produce more software. Coding is no longer its biggest bottleneck, the company said. Instead, source control, builds and automated testing infrastructure designed around human development speeds are struggling to keep up.

Anthropic described its workload as unusually large. Its main branch history alone approaches 100 GB, while a compressed working copy exceeds 25 GB. Automated testing jobs perform roughly 25 million code checkouts a day, before accounting for developer environments, agent sandboxes and research workloads.

The company had used partial copies, repository snapshots and several hundred internal Git mirrors to manage that demand. Those workarounds became harder to operate as the repository and merge rate grew.

Scion takes a different approach by assuming its users, mostly agents and automated jobs, remain online. Servers handle operations such as creating commits, merging changes and retrieving file history, while clients fetch only the data they need.

That reduces the need to maintain large local repository copies. Some tools can query or modify code entirely through an API, while jobs that require files can retrieve selected directories or load files on demand.

Anthropic said Scion’s CPU and memory requirements are 200 times smaller than those of its previous Git mirroring infrastructure. Code checkouts for testing and research jobs are about 100 times faster, taking seconds rather than minutes.

The migration also freed up 25% of CPU capacity used by its continuous integration workloads, according to the company.

Claude played a central role in building the infrastructure needed to accommodate its own output. Anthropic said an early internal version of Claude Mythos Preview rewrote a prototype ingestion pipeline to use 20 times less memory and run four times faster.

Engineers remained responsible for architecture, design tradeoffs and enforcing system constraints. The company said agents sometimes introduced shortcuts or unnecessary complexity, requiring human intervention.

Team members still review every change request for Scion, although Anthropic said their focus has shifted toward the intent of changes rather than reading every line of code.

The company also uses Claude agents to monitor individual infrastructure components. These agents examine telemetry and logs for unusual behavior alongside conventional alerts.

The growing volume of changes has prompted Anthropic to reconsider other development practices. It said hundreds or thousands of changes can merge while tests are running, complicating traditional merge queues. The company has instead invested in testing after changes merge and before deployment.

Anthropic cautioned that Scion is suited to enormous repositories and large numbers of coding agents. Adopting it would require extensive changes across an organization’s development tools, and the company is not presenting it as a general replacement for Git.

It expects similar infrastructure pressures to become more common over the next six to 12 months. Rather than offer a broadly available version of Scion, Anthropic said it wants to share its approach with GitHub, GitLab and the open source community.