Anthropic went from 2,500 code merges a day to over 30,000, so it took its monorepo off git

Anthropic averaged around 2,500 code changes merged per day in May 2026 and merged over 30,000 a day by September, and the source control underneath it gave way. Anthropic says it built its own version control system, called Scion, and moved its monorepo off git in five months.

What Anthropic was running before Scion, which at its peak included several hundred internal git mirrors. Anthropic captions it "The typical setup for scaling git checkouts." Image: Anthropic

The scale is the reason it bothered. The history of the main branch alone is "almost 100 GB" and a compressed working copy is over 25 GB, Anthropic says, and it runs "on the order of 25 million code checkouts per day for CI, before counting developer environments, agent sandboxes, and research jobs". Fetched naively, it reckons that would be over 600 PB of git traffic a day.

That last figure is a hypothetical rather than traffic Anthropic served, and the 30,000 counts merges into one internal monorepo, not lines, commits or pull requests.

Scion's checkout architecture in Anthropic's own diagram, where a job can ask the server for a diff rather than check code out. Image: Anthropic

Several hundred git mirrors bought time, not a fix

Anthropic went through the standard remedies first: shallow clones, sparse checkouts, promisor remotes, internal git mirrors and repository snapshots kept in blob storage. "We've used all of them," it says, "at our peak, we operated several hundred internal git mirrors to deal with the load."

Each of those got harder to run as the merge rate climbed. Snapshots fell further behind the tip of the main branch between refreshes, Anthropic says, and keeping history-dependent operations such as diff and blame working on partial clones took more and more care.

The underlying mismatch is a design assumption. Git expects a client to hold a complete copy of the repository and work offline, which Anthropic calls "the right design" for most repositories and a large part of why git became the standard, and which stops fitting when nearly every client is an agent or a CI job.

How Scion holds a change, in the figure Anthropic captions "The shape of a changeset in Scion." Image: Anthropic

Scion does the commit on the server

Anthropic lists five principles behind Scion, and the one a git user would feel first is where work lands. "In Scion, every commit, rebase, or merge happens on the server, so there's nothing to push," it says, with the cost of a commit scaling to the size of the change rather than the size of the repository.

The repository is never assembled in one place either. Scion's data is distributed and each component scales on its own, and Anthropic says "there is no fully materialized copy of the repo anywhere, even on the server", with no equivalent of git's packfiles and therefore "no repository maintenance".

A working copy is optional and sparse by design, so a CI job that wants the diff between a commit and its merge base can ask the server for it instead of checking anything out. Branches and tags are discouraged outside the main ref and release branches, because Anthropic treats state that every server has to agree on as a scaling problem in waiting.

None of this is unprecedented. Anthropic names Google's Piper and Meta's Sapling as the prior art it built against, and says "Scion is by no means a suitable replacement for git for general use".

Weekly merges into Scion and Antique themselves, the two systems Anthropic built while it was migrating. Image: Anthropic

Claude rewrote the pipeline that fed Scion

The migration needed a pipeline to pull git history into Scion, and Anthropic largely handwrote the first version. It then asked "an early internal snapshot of Claude Mythos Preview" to rewrite the whole thing, and says that "with no further instructions, it produced a version that consumed 20x less memory and was 4x faster, while preserving all of the implicit invariants around ingestion ordering and data structures".

That is one rewrite of one pipeline with no benchmark method published, which puts it alongside Anthropic's other self-run numbers: it also scored Opus 5.5 on Terminal-Bench itself rather than going through the public board.

Scion was not the only thing built in those five months. Anthropic says it also shipped Antique, a code review surface, plus code search and a source control IAM system, and it publishes no figure for what the migration cost, how many tokens it consumed, or how much of Scion's code Claude wrote.

Nobody outside Anthropic gets to use any of it. Anthropic says it would be "difficult to offer Scion in a form that would be broadly useful to other organizations", because it "diverges significantly from git" and brings its own code review, code search and CI with it.

The concepts, Anthropic adds, "really only make sense for truly enormous monorepos, with either tens of thousands of engineers or swarms of agents prolifically writing code". It sells the tooling for the second of those: Claude can now write its own programs to run up to 1,000 agents and start them itself.