On October 13, Anthropic published an engineering retrospective disclosing that it spent five months moving its entire monorepo from git to Scion, a version control system it built in-house. In the same period it also built a code review tool called Antique. The core team was just seven people.
The trigger was merge volume. In May 2026, Anthropic's monorepo averaged roughly 2,500 merges a day. By September that figure had passed 30,000. The write-up puts it this way:
“As our merges per day grew more than tenfold, source control became our biggest development bottleneck.”
Writing code was no longer the constraint. What was holding things up was source control, builds and CI.
Where git stops scaling
Anthropic listed several numbers. Trunk history is close to 100 GB, a compressed working copy is more than 25 GB, and CI performs about 25 million checkouts a day. If each one pulled everything in full, git traffic would exceed 600 PB per day. At peak the company ran hundreds of internal git mirrors, and as merges sped up, the mirrors and their prebuilt snapshots became harder and harder to maintain.
The root cause, according to the write-up, is a mismatch of design goals. Git was built for offline, distributed collaboration between humans, with everyone holding a full copy of the repository. Anthropic's main clients today are agents and automated jobs that are always online.
Scion therefore inverts the model. History, blame, commits, merges and rebases all live on the server, while the client keeps only a limited local cache and working copy and fetches data on demand. A commit completes on the server, and its cost scales with the size of the change, not the size of the repository. Only trunk and release branches are kept, and moving a branch pointer is a compare-and-swap on a single database row. Clients go through a local daemon, then tiered caches, and only then the server, with hit rates above 99% at each cache layer.
Doing the math
A back-of-envelope check on the published figures: 600 PB divided by 25 million checkouts is about 24 GB per checkout, which lines up with the “more than 25 GB” working copy. That suggests the 600 PB number is an upper bound estimated on the assumption of a full pull every time.
Spread over 24 hours, 30,000 merges works out to a little over 20 a minute. Under git's push model, every push has to sync with the remote first, and at that rate conflict retries would clog the queue. Scion moves merging to the server and removes the push step. Anthropic also dropped its global merge lock, relying instead on automated tests that run after merge and before deployment, and when concurrent changes collide, Claude is asked to fix them.
By the company's own account, the payoff is this: Scion's infrastructure uses about 200 times less CPU and memory than the old git mirror fleet, checkouts for CI and research jobs dropped from minutes to seconds, and CI alone freed up roughly 25% of CPU capacity.
What Claude did along the way
The migration itself used models too. An early, hand-written git-to-Scion import pipeline was rewritten by Claude Mythos Preview, cutting memory use to one-twentieth and running about four times faster. The write-up says Claude also found some algorithm-level optimizations and wrote large volumes of tests, simulations and fuzz tests. The engineers' job became holding the design intent, rejecting unreasonable approaches and reviewing what a change is meant to do rather than reading the code line by line.
The retrospective does not go deep on Antique. It says only that it lets teams set semantic rules for approvals, separating commits that touch only comments from those that change sensitive logic so human reviewers can concentrate on the latter.
Scion will not be open-sourced. Anthropic is blunt about it: “Scion is by no means a suitable replacement for git for general use.” Its internal components are too heavily customized to package for others. The company hopes GitHub, GitLab and the open-source community will borrow ideas such as server-side RPC and sparse checkout to improve git. Google's Piper and Meta's open-source Sapling went down a similar road; the difference this time is that the pressure to rewrite comes from the volume of code written by AI.
Sources: CocoLoop, Anthropic Engineering Blog. Daily merge counts, checkout counts and resource-reduction multiples are as reported by Anthropic; the per-checkout size is our own rough calculation from its figures.