Anthropic Dropped git for Scion in Five Months: 30 000 Merges a Day

In May 2026, Anthropic was merging about 2 500 code changes a day into its monorepo. By September that figure had grown to 30 000+. Writing code is no longer the problem: Claude iterates on research loops, optimizes complex programs and spins up services from scratch. The bottleneck has moved to infrastructure built for humans: version control, builds and CI. To clear it, the company migrated its monorepo from git to its own system, Scion, in five months. Alongside it, Anthropic built Antique, a code review tool, plus code search and an IAM system for source control.

Where git stopped fitting

Git is decentralized: every clone is a full repository that can compute diffs, read history and work without a network. For most projects, that is the right design. At Anthropic, it started to get in the way. The history of the main branch alone takes up almost 100 GB, and a compressed working copy is over 25 GB. CI alone runs about 25 million checkouts a day, and that excludes developer environments, agent sandboxes and research jobs. A naive download approach would have meant over 600 PB of git traffic every day.

The team tried shallow clones, sparse checkout, promisor remotes, internal mirrors and snapshots in blob storage. At peak, several hundred git mirrors were running. They bought time, but as the merge rate climbed they got harder to maintain. Snapshots lagged behind the tip of main, and diff and blame on partial clones demanded more and more upkeep. The team's conclusion: git was being used as a centralized system with a single authoritative copy, and its clients were agents and jobs that are online by definition. The properties that matter for offline work were holding back scaling.

How Scion works

The unit of work in Scion is the changeset, a record in a write-ahead log. Instead of a pointer to a tree, it stores a list of files changed relative to the first parent, plus content-addressed references to compressed data in blob storage. A changeset is created on the server and is immediately available everywhere. There are no packfiles, no locks and no cross-server coordination. A ref, when one is needed, moves as a compare-and-swap on a single row in Spanner, separately from changeset creation.

Questions like "who last changed this line" are answered by derived data: structures for blame, path histories and commit-graph skip lists. Each type is defined recursively through values for the parents. It can be computed eagerly or lazily, and its internals can change without touching the client.

The scn client does not tell the server what it already has. It starts from the root tree, requests the missing trees, then the blobs. Every request targets an immutable object, so any cache can serve it. A node talks to a local scion-daemon, which pulls data from regional scion-cache instances, which in turn go to the authoritative scion-server on a miss. The strict separation of tiers isolates failures: a regional cache outage should not take down the core. Sparse and virtualized checkouts are supported without special logic, because the client simply doesn't ask for trees it doesn't need.

What the company lists as results:

• growth from 2 500 to 30 000+ merges a day (May–September 2026);
• monorepo migration off git in five months;
• Scion infrastructure 200 times smaller than the git mirrors in CPU and memory;
• checkouts in CI and research jobs two orders of magnitude faster: seconds instead of minutes;
• cache hit rate above 99% at every tier;
• 25% more processing capacity, in CI alone.

What to take from this

Anthropic's principles draw on Google Piper and Meta Sapling: clients are always online, the server does the heavy lifting, commits are cheap and lock-free, and a working copy is optional. Every client request is also available as an API, so CI jobs often don't need a local repository at all. Internal pipelines for automated bisect and revert rely on this; at large volumes they would otherwise be impractical.

Anthropic itself notes that its workload is unusual and that many teams are unlikely to hit these problems any time soon. But if a team is already running dozens of agents, it is worth counting how many checkouts it does per day, what the mirrors cost, and how far snapshots drift from the tip of main. A starting point is moving metadata such as diff, merge base and blame into a service, as Scion did early in the migration.

Anthropic also says that Claude Mythos Preview and Claude Tag, which appeared inside the company in the same months, increased the need for Scion and sped up its development.