Anthropic Replaces Git with In-House Version Control System Scion for AI Development

On Oct. 14, Anthropic announced it has replaced Git with Scion, a custom version control system built from scratch by a seven-person team over five months with heavy assistance from Claude. The move came as Claude's coding speed turned Git into an internal bottleneck.

Anthropic said daily merges into its monorepo rose from about 2,500 in May 2026 to roughly 30,000 in September 2026, or an average of one merge every 2.8 seconds, 24 hours a day. Git, designed for offline work with full local copies, required nearly 100GB of history to be cloned onto each machine. At peak, Anthropic ran hundreds of Git mirror servers to handle the traffic.

Scion supports virtualized checkouts, pulling only the folders an AI agent is working on and streaming file data from cache on demand. It removes local packing and global locks, so commit time depends only on the amount of changed code. Compared with the previous Git mirror setup, Scion cuts CPU and memory consumption by a factor of about 200 and maintains cache hit rates above 99 percent. Code fetching for AI and test tasks dropped from minutes to seconds, saving 25 percent of compute for the company's continuous integration systems. The team also built a companion review tool called Antique, along with code search and permission systems.

Claude was deeply involved throughout development: an internal early model rewrote the engineers' data migration pipeline, cutting memory usage to 1/20 and making the pipeline four times faster. Anthropic said it does not plan to sell Scion commercially, and it is publishing the architecture because software engineering tools designed for human-speed workflows become obstacles when large numbers of AI agents generate, test, and fix code in parallel. The company expects other organizations to hit similar bottlenecks within six to twelve months, roughly from April to October 2027.