Anthropic Builds In-House Version Control System Scion, Migrates Entire Codebase Off Git in Five Months

What happened

On October 13, Anthropic published “Redesigning source control at Anthropic” on its official engineering blog, announcing that in-house version control system Scion is live and the company has fully migrated off Git. A code review tool called Antique shipped alongside. The migration was done by a seven-person team in five months, during which the company's daily merged-code volume grew tenfold.

According to the blog, Claude's ability to write code was both what put the company in this situation and how the team was able to build Scion quickly: an early internal Claude Mythos Preview snapshot rewrote the git-to-Scion ingestion pipeline, cutting memory use to one-twentieth and running four times faster, while preserving all implicit data-structure invariants.

Key facts

1. Scale numbers: The main branch history alone is nearly 100GB, a compressed working copy over 25GB; CI performs about 25 million code checkouts per day; fetched naively with Git, that would mean over 600PB of traffic per day.
2. Design principles: Scion is designed around “agents are always online”: client-server asymmetry (commits, merges and rebases all happen server-side), everything horizontally scalable, cheap incremental lock-free commits, and working copies optional and sparse by default. The design draws on lessons from Google's Piper and Meta's Sapling.
3. Migration results: Scion's infrastructure footprint is 1/200th of the previous Git mirroring setup in CPU and memory; checkouts went from minutes to seconds, two orders of magnitude faster; CI realised a 25% CPU capacity saving.
4. No general product: Anthropic says Scion is only its own answer and will not become a general Git replacement; what it is sharing is the design thinking, so the industry can apply Scion's successful parts to Git, and it wants to work with GitHub, GitLab and the open-source community on developers' changing needs.

Background

This is the first time a top AI lab has publicly abandoned Git because of agent-written code. Large-scale monorepo version control previously had internal solutions such as Google's Piper and Meta's Sapling. Anthropic expects these problems to become more common at other companies using AI to write code over the next six to twelve months.