Anthropic threw out its version control system and built a new one. On October 13 the engineering team published how it moved its monorepo off Git in five months, onto an in-house system called Scion. One number explains why: in May 2026 the company merged around 2,500 code changes a day. By September it was merging more than 30,000.
Seven people did this. In the same five months they also shipped a code review tool called Antique, code search, and an IAM system for source control.
Git rests on an assumption that no longer holds at Anthropic
Git assumes every client might work offline and therefore needs a complete copy of the repository. At Anthropic, almost every client is an agent or a job, and those are online by definition. The main branch history is close to 100 GB; a compressed working copy is over 25 GB. CI alone pulls roughly 25 million checkouts per day. Fetched naively, that would be more than 600 petabytes of Git traffic daily.
Anthropic tried all the usual escape hatches first: shallow clones, sparse checkouts, snapshots in blob storage, and at peak several hundred internal Git mirrors. That bought time and got worse as the merge rate climbed.
Scion flips the roles. The server answers questions about file history and blame, creates commits, performs merges and rebases. The client keeps a bounded local object cache and fetches what it needs when it needs it. Commits happen server-side, so there is nothing to push. Branches and tags are discouraged, because global state that every server has to agree on eventually becomes the bottleneck.
What the rebuild bought
The checkout infrastructure now needs 200x less CPU and memory than the Git mirrors did. CI and research jobs get code in seconds instead of minutes, which freed up 25 percent of CI’s compute capacity on its own. Every cache tier holds a hit rate above 99 percent.
The engineers’ real job was saying no
Anthropic writes that every engineer on the project would now rate Claude as a better coder than themselves. One example: the Git-to-Scion ingestion pipeline had been written by hand and the team thought it was good. Then they got an early internal snapshot of Claude Mythos Preview, asked it to rewrite the whole thing, and the new version used 20x less memory and ran 4x faster.
The team still describes the human role differently, as a brake. Agents repeatedly violated core system constraints to satisfy a pattern that should never have been allowed in the first place. Several systems brute-forced the question of when a change merged to main by reading the commit log linearly instead of asking the code review tool. And the company’s heavy use of adversarial review pushed Claude toward over-engineering parts of the system against hypothetical concerns.
Anthropic is not releasing Scion. It doesn’t suit normal repositories, and the team thinks a public version built to please everyone wouldn’t be much use. What it offers instead is the reasoning: a Git server doesn’t have to work the way a Git client does, and comprehensive server-side RPCs along Scion’s lines could be built into existing platforms. Anthropic expects this class of problem to show up more widely within six to twelve months.