Anthropic Replaces Git With Scion After Claude Code Surge

Anthropic has retired Git across its primary monorepo, deploying a custom version control system named Scion to handle automated software development driven by its Claude AI models. A team of seven engineers built and migrated to Scion along with a code review tool called Antique in roughly five months. The shift was triggered when daily merges jumped twelvefold, rising from around 2,500 in May to more than 30,000 by September, as Claude began writing the vast majority of internal codebase changes.

Git's traditional model of downloading complete local repositories threatened to generate over 600 PB of daily network traffic across Anthropic's continuous integration fleet. Scion resolves this by turning source control into a server-centric service based on RPC calls, virtualized checkouts, changesets, and backend storage powered by Google Spanner and cloud object storage. In benchmarks, the new deployment consumes just 1/200th of the CPU and memory of the previous Git mirror fleet while maintaining cache hit rates above 99% across tiers. CI and research checkouts dropped from minutes to seconds, freeing up 25% of the fleet's total CPU capacity.

Claude itself authored most of Scion's codebase under human oversight. During development, an early internal build of Claude Mythos Preview refactored a Git-to-Scion ingestion pipeline without explicit instruction, cutting memory utilization by a factor of 20 and multiplying execution speed by four. Anthropic also deployed "superfans"—dedicated Claude instances monitoring telemetry logs for scion-daemon, scion-cache, and Antique to flag subtle errors that rule-based alerts miss.

For software practitioners, Anthropic's transition highlights how agentic coding disrupts conventional repository design and merge queues. While Anthropic has no immediate plans to open-source Scion, the architectural lessons suggest high-concurrency environments must move away from repository-wide locks and offline clones. Teams scaling AI agents will increasingly need server-side RPC architectures, lazy file fetching, and post-merge test validation paired with automated conflict resolution.