Redesigning source control at Anthropic
Short summary
Anthropic built Scion, a custom version control system, to replace Git because Claude's code generation capability pushed their merge rate from 2,500 to 30,000+ code changes per day—a scale Git's distributed design couldn't handle. The system treats clients as always-online agents and moves computation to the server, ditching Git's model of complete local repositories.
- Anthropic's monorepo became unmanageable under Git: the history alone is 100 GB, and they were doing 25 million checkouts daily for CI, which would generate 600 PB of traffic if done naively.
- Scion's core design flips Git's assumptions: clients stay thin and online, the server handles all history queries and merges, and there's no single fully-materialized copy of the repo anywhere—everything is distributed and horizontally scalable.
- The system uses "changesets" (write-ahead log entries) instead of Git commits, storing only what changed relative to the parent plus content-addressed pointers to blob storage, allowing high-throughput acceptance with no packfiles or cross-server locking needed.
Full summary
Anthropic's code generation has grown so fast that git became the limiting factor. They went from merging 2,500 code changes daily in May 2026 to over 30,000 by September—a twelve-fold increase in four months. At that scale, their monorepo history alone hit 100 GB, and they were performing 25 million checkouts per day for CI jobs. Naive git operations would have required 600 petabytes of daily traffic. They tried standard workarounds: shallow clones, sparse checkouts, and eventually operated several hundred internal git mirrors. But these mitigations couldn't keep pace as their merge rate kept climbing, and operations like diff and blame started breaking on partial clones.
The core problem was that git was designed for distributed, offline work—every developer gets a complete copy of the repository. Anthropic's reality is the opposite: almost all their clients are agents and jobs that are always online. So they built Scion, their own version control system, and migrated their entire monorepo in five months. The design flips git's assumptions. Clients always reach a server instead of maintaining complete local copies. The server handles history queries, blame, commits, merges, and rebases. Clients keep only a bounded cache and sparse working directories. There's no fully materialized copy of the repo anywhere, even on the server side.
Scion's architecture reflects this shift. Commits happen server-side and are cheap, incremental, and lock-free—the cost scales with change size, not total repo size. The system exposes all source control queries via API, so CI jobs and internal tools can ask questions about history without cloning anything. Working copies are optional and sparse by default; a CI job needing a diff can request it from the server rather than checking out files. Everything scales horizontally with no central bottleneck. While Anthropic's workload is extreme, they expect other companies using AI coding tools will hit similar constraints soon, which is why they're sharing what they learned.