Anthropic Migrates Monorepo to In House Source Control System

Problem and Motivation

Anthropic's monorepo grew rapidly as AI assisted coding increased. Code merges per day rose from roughly 2500 in May 2026 to over 30000 by September 2026. Git became a bottleneck for this scale. The main branch history alone is over 100 gigabytes and a compressed working copy is over 25 gigabytes. The system performs about 25 million code checkouts per day for CI. Naively fetched that would be over 600 petabytes of git traffic daily. Internal mirrors and snapshot workarounds became difficult to operate as merge rates increased.

Scion Design and Architecture

Anthropic built in house version control system and fully migrated its monorepo from git in five months. Scion was designed around principles suited for always on clients. Clients do not need complete local copies since they are typically online. The server handles history blame merges and rebases while clients maintain a bounded cache. Changesets store file changes and content addressed blob pointers enabling high throughput. Derived data structures such as file blame and path history are computed recursively from changesets. The system avoids packfiles and per repository materialized copies achieving 200x smaller infrastructure footprint than previous git mirroring. CI and research jobs check out code two orders of magnitude faster in seconds rather than minutes.

Lessons and Impact

A team of seven engineers built Scion and the Antique code review tool in five months despite a tenfold increase in daily merges. Claude helped rewrite ingestion pipelines achieving up to 20x memory reduction and 4x speedup. Engineers shifted role to high level design and system constraints frequently telling Claude no to prevent over engineering. The system supports rich directory level access control and semantic code review constraints. Operational excellence is enhanced by component monitoring superfans that track telemetry and logs. New workflows favor post merge automated testing over global merge queues for agent scale.

Sources: Techmeme · Techmeme