Anthropic Builds Custom Source‑Control System to Handle Claude‑Generated Code

Anthropic, the company behind the Claude language model, says its AI writes software at a speed and scale it claims it hasn’t seen before, producing thousands of code changes each day. In May 2026 the company averaged about 2,500 code changes merged per day, and by September that number grew to more than 30,000 per day.

Git, which works well for most projects, struggles with a monorepo of that size. The history of Anthropic’s main branch alone is nearly 100 GB, and a compressed working copy exceeds 25 GB. Daily CI jobs check out code tens of millions of times, which would generate huge network traffic if Git were used naïvely.

To address the bottleneck, Anthropic built a custom source‑control system called Scion and migrated its entire monorepo from Git in five months.

Scion’s design assumes that all agents are online and that the server does most of the work. Clients keep only a small cache of needed objects and ask the server for file history, blame, merges, and other metadata via API, eliminating the need for each client to hold a full copy of the repository.

In Scion, the basic unit of work is a changeset, an entry in a write‑ahead log. A changeset stores the list of files that changed relative to its first parent, along with content‑addressed pointers to the compressed file contents in blob storage, rather than a pointer to a tree as in a Git commit.

Scion also builds derived data structures such as blame and path history on top of changesets. These structures are computed recursively from parent changesets, allowing the system to answer queries quickly without storing all information in each changeset.

Because the server exposes all source‑control queries via API, many internal tools no longer need local repositories. Continuous‑integration jobs can request diffs, merge bases, and other metadata directly from the server, and sparse or virtualized checkouts mean that jobs only pull the directories they actually need.

Anthropic says Scion is a bespoke solution that will not be released as open source. The company believes the design could be useful for other organizations that face similar scaling challenges from AI‑generated code.

As large language models become more common, the need for source‑control systems that can handle thousands of daily changes may become a widespread concern.