Anthropic replaces Git internally as daily code merges top 30,000

Anthropic’s coding workload has outgrown the system it used to manage code changes. In an October 13, 2026 engineering account, the company disclosed a five-month migration from Git to Scion, its own version-control system. Anthropic says Claude-driven development made source control—not writing code—its biggest development bottleneck as daily merges climbed above 30,000.

The workload that exhausted the workarounds

The pressure built quickly. Anthropic averaged about 2,500 merged code changes a day in May 2026; by September, that exceeded 30,000. The company attributes the strain to Claude’s ability to iterate on research, optimize programs and build services at a scale its developer infrastructure was not designed to support.

Its monorepo—a single repository holding its code—had almost 100 GB of main-branch history and a compressed working copy exceeding 25 GB. Continuous integration, or CI, the automated checking stage of development, required about 25 million code checkouts daily. That count excluded developer environments, agent sandboxes and research jobs.

Anthropic had already tried partial downloads, repository snapshots and several hundred internal Git mirrors. Those measures bought time, but snapshots slipped further behind the latest code between refreshes. Maintaining history-dependent operations on partial repositories also became harder. The mirror fleet strained capacity and stopped scaling when the repository exceeded individual mirrors’ disk or memory limits.

Scion moves the work onto servers

The redesign began with a different assumption about who—or what—uses source control. Git supports self-contained repositories that can inspect history and check out revisions offline. Anthropic was already using one authoritative repository, while nearly all its clients were online agents and jobs. Scion therefore assumes a reachable server rather than an independently capable local copy.

Servers create commits, perform merges and rebases, and answer questions about file history and who changed a line. The client keeps a bounded local cache, puts working files on disk and watches for file changes. Repository data is distributed across independently scalable components; even the server does not hold a complete, fully materialized copy.

Scion records each change as a server-created “changeset”: a list of files changed relative to its first parent, with pointers to compressed file contents. Creating one does not require accepting Git’s compressed object bundles, locking across servers or moving a shared branch pointer. Separate derived indexes make history and line-attribution queries efficient without burdening the client.

For many tools, checking out code becomes unnecessary. Scion exposes its source-control queries through APIs, interfaces that let software request operations directly. A CI job can ask for the difference between revisions without maintaining a local repository. When files are needed, sparse checkouts retrieve selected directories, while virtualized checkouts materialize files only on demand.

Requests for missing directory and file objects pass through a cache on the compute node, then regional caches, then authoritative servers. Objects never change after storage, so any cache holding one can answer. Regional caches stream data in the format clients consume, avoiding the work of calculating customized compressed transfers for each request.

Smaller infrastructure, with a narrower design target

Anthropic reports that Scion’s CPU and memory requirements are one two-hundredth of its former Git-mirroring infrastructure. CI and research checkouts are 100 times faster, completing in seconds rather than minutes; CI gained 25% in CPU capacity. Each cache tier has a reported hit rate above 99%, and clients and caches no longer need repository maintenance.

The five-month effort also delivered Antique, an internal code-review surface, alongside code search and source-control access management. The caching architecture explicitly separates failure domains: Anthropic enforces the tiers to prevent a regional cache failure from overwhelming upstream servers and taking down the whole system.

Anthropic does not present its workload as typical. It says Git’s decentralized design remains right for the vast majority of repositories, and it does not expect many organizations to face its problems in the very near term. Its expectation is that wider use of AI coding tools will eventually bring similar pressure points elsewhere.