Anthropic says it has moved its main monorepo from Git to an in-house version control system called Scion after code merges grew beyond the workflow built for the company’s earlier scale. The company reported about 2,500 merged changes a day in May 2026 and more than 30,000 a day by September, a rise of more than tenfold. It says the full migration took five months.

The post is a description of Anthropic’s internal system, not an announcement that Git is generally obsolete. Anthropic says the design fits unusually large repositories and workloads dominated by always-online agents and jobs. Most software teams, it says, are not yet facing the same constraints.

Why Git became a bottleneck at Anthropic

Anthropic’s account focuses on the size and traffic of its monorepo. The main branch history is nearly 100 gigabytes, and a compressed working copy exceeds 25 gigabytes. Continuous-integration jobs perform about 25 million checkouts per day, before counting research jobs, agent sandboxes and developer environments. The company estimates that fetching full copies naively would produce more than 600 petabytes of daily traffic.

Git’s distributed design gives each clone a complete repository that can work offline and inspect history locally. Anthropic says that remains a strong fit for most projects. Its own environment, however, increasingly used Git as a centralized service: most clients were online jobs asking for data from one authoritative repository.

The company had already tried sparse checkouts, shallow clones, mirrors and stored snapshots. At peak it operated several hundred internal Git mirrors. Anthropic says the mirrors became harder to maintain as merge volume rose, while snapshots drifted from the main branch and partial clones made history-dependent operations such as blame more difficult.

What Scion changes

Scion makes the server responsible for creating commits, merges and rebases and answering questions about history and blame. Clients keep a bounded local object cache and request only the data they need. Anthropic also built related tools, including Antique for code review, code search and a source-control permissions system.

The system stores changes as server-created changesets and keeps file contents in content-addressed storage. Anthropic says that avoids large packfiles and global reference updates. It can compute history, diffs or merge bases through an API, so some jobs can ask questions without checking out the repository at all.

When a working copy is needed, Scion supports sparse and virtualized checkouts. A client can materialize only the requested directories or files rather than downloading a complete repository. The company says its cache tiers have hit rates above 99 percent, its infrastructure uses about 200 times less CPU and memory than its former Git-mirroring setup, and CI and research checkouts run roughly 100 times faster. It also reports a 25 percent reduction in CI CPU use.

A specialized migration, not a Git replacement

Anthropic says a team of seven built Scion, Antique and related systems in five months while merge volume rose. The company credits Claude with helping write the software, but says engineers made the system-level design decisions and constrained agents when generated changes conflicted with core requirements.

The trade-off is that Scion depends on a reachable server and requires a broader migration of code review, search and CI. It is not designed to preserve Git’s full offline clone model. Anthropic says it would be difficult for other organizations to adopt because its internal tools are bespoke and the system diverges substantially from Git.

The practical point is narrower than the headline numbers: Anthropic redesigned its own source-control workflow around millions of daily checkouts and a much higher merge rate. The company presents Scion as one response to agent-scale development, not a reason for ordinary repositories to abandon Git.