Anthropic moved its monorepo from Git to an in-house system, Scion

On 13 October 2026, Anthropic published Redesigning source control at Anthropic. The post says the company fully migrated its monorepo from Git to an in-house version control system, Scion, in five months, and that it built a code review surface called Antique in the same stretch. Simon Willison’s same-day note quotes the merge figures and the seven-person, five-month account. Every measurement below is Anthropic’s.

How fast the monorepo grew

Anthropic writes that in May 2026 it averaged around 2,500 code changes merged per day, and that as of September it merged over 30,000 per day. The opening line calls that rise more than tenfold. The post says the history of the main branch alone is almost 100 GB, a compressed working copy is over 25 GB, and the company performs on the order of 25 million code checkouts per day for CI, before developer environments, agent sandboxes, and research jobs. Fetched naively, Anthropic says, that volume would be over 600 PB of Git traffic per day. The 600 PB line is a hypothetical full fetch, not a measured traffic total.

At the peak of the Git setup, Anthropic says it ran several hundred internal Git mirrors, after shallow clones, sparse checkouts, promisor remotes, and blob-storage seeds.

What Scion is

Scion assumes clients are always online. Anthropic contrasts that with Git, where every clone can compute diffs, history, and blame offline from a complete copy. In Scion the server answers history and blame, creates commits, and performs merges and rebases. The client keeps a bounded local cache, lays out the working copy, and watches the filesystem. The working copy is optional. When a job does need files, the protocol is built around sparse checkouts, which contain only the directories a client asks for, and virtualized checkouts, which materialize files on demand. A CI job that only needs a diff can ask the server and skip the checkout.

Anthropic names Google’s Piper and Meta’s Sapling as prior art. The unit of work is a changeset: a write-ahead-log entry that stores the files changed relative to the first parent, plus content-addressed pointers to the compressed contents. Blame, path history, and other derived data are computed from those changesets. The post says the data is distributed, there is no fully materialized copy even on the server, and the authoritative store is Spanner plus cloud storage. Cache servers and the authoritative servers answer the same RPCs.

What Anthropic says it gained

Anthropic says the Scion footprint is now 200x smaller than the previous Git mirroring infrastructure, in CPU and memory. It says CI and research jobs check out code two orders of magnitude faster, in seconds rather than minutes, and that CI alone realized a 25% CPU capacity win. It also says each cache tier keeps a hit rate over 99%. A team of seven people built Scion and Antique, plus related systems, and finished the migration in five months, while the volume of code merged per day increased by 10x. Those figures are Anthropic’s account of its own fleet.

The post also says an early internal snapshot of Claude Mythos Preview rewrote a handwritten Git-to-Scion ingestion pipeline so that it used 20x less memory and ran 4x faster. That comparison is Anthropic’s account of one internal rewrite. It was not rerun here.

What Anthropic says Scion is not

Anthropic writes that Scion “is by no means a suitable replacement for git for general use.” It says the design fits enormous monorepos, with tens of thousands of engineers or swarms of agents writing code, and that a public version trying to please everyone would not be very useful. The post does not announce an open-source release, a download, or a hosted product. It says the useful step for everyone else is to apply parts of the idea to Git: a Git server does not have to work like a Git client, forges can offer server-side RPCs in a form a Git client could learn, and Git’s sparse and shallow features can be made to compose. It names GitHub, GitLab, and the open-source Git community as the partners it has in mind.

Willison’s note quotes Scion engineer Nick G, in a Hacker News reply this account did not open: “Since Claude is by far the heaviest user of Scion and is necessarily online, this just wasn’t important to do.” That sentence is in Willison’s post. The thread itself was not opened.

What to take from it

The counts to use are the ones in the 13 October post: around 2,500 merges a day in May, over 30,000 in September, seven people, five months, on the order of 25 million CI checkouts a day, a 200x smaller mirror footprint, checkouts in seconds, and a 25% CI CPU saving. All of those are Anthropic’s measurements. There is nothing to install. The request to other teams is that forges add server-side queries and sparse protocols that compose.