Anthropic builds Git replacement Scion with 7 engineers in 5 months

Anthropic did something that seems unreasonable by traditional engineering intuition: with 7 engineers over 5 months, it developed from scratch an internal source control system called Scion and a companion code review tool called Antique, and fully migrated its entire monolithic code repository away from Git. In the past, this team size and time window would usually only be enough to complete an early prototype of a version control system.

The driving force was not technical idealism, but load. Data published by Anthropic shows that its average daily merged code changes grew from about 2,500 in May this year to over 30,000 in September, a more than 12-fold increase within 4 months, and it expects this growth not to slow down in the short term. The potential demand for code merging has almost no upper limit—whenever one bottleneck is solved, code output increases further.

Anthropic itself judges that not many companies will encounter the same AI programming scale bottleneck in the short term, but it expects other companies to run into similar problems before long. This is also why they decided to publicly share Scion's design and development experience. In other words, this system is more like a sample that arrived early, rather than a solution that all teams must copy right now.

Git is a decentralized system. Every clone gets a complete, independent code repository, and even fully offline it can compute diffs, view history, trace who last modified a line of code, and check out any version. For the vast majority of repositories, this design is very reasonable, and it is an important reason Git became the industry standard. But Anthropic's repository size has already crossed beyond this design's comfort zone.

The specific numbers are: the main branch history alone is close to 100GB, and the compressed working copy exceeds 25GB. The CI system alone performs about 25 million code checkouts per day, not counting development environments, Agent sandboxes, and research tasks. If the needed data were fetched directly from Git every time, the daily data transfer volume would exceed 600PB.

Anthropic did not fail to try conventional solutions. They used shallow clones, sparse checkouts, Promisor Remote, and also built internal Git mirrors that stored repository snapshots in object storage. At peak load, they ran several hundred internal Git mirror servers. These solutions did hold up for a while, but they became increasingly difficult to maintain as merge volume grew: between two snapshot refreshes, the gap between the snapshot and the latest version of the main branch grows larger and larger; and to make partially cloned repositories correctly perform history-dependent operations such as diff and blame, the maintenance investment also grows higher and higher.

After re-examining usage patterns, Anthropic found that they had in fact always used Git as a centralized system. All development work revolved around one authoritative repository, and the clients were almost all Agents and automation tasks that were always online anyway. The features Git provides for distributed and offline development instead limited scalability as the repository kept growing.