Anthropic Scion: Why Claude-Scale Merges Pushed Out Git

The headline doing the rounds is "Anthropic ditched git". The engineering post behind it, published on 13 October 2026, says something narrower and more interesting: one company's monorepo outgrew git's design once Claude was writing most of the code, so seven engineers built Anthropic Scion, a server-centric system, and moved the repository onto it. Git is still git. GitHub is not mentioned as something Anthropic left. And Scion is not for sale.

Key points

• Merges into Anthropic's monorepo went from about 2,500 a day in May 2026 to more than 30,000 a day by September, according to the post.
• Scion and a code review tool called Antique were built by a team of seven, and the monorepo was fully migrated off git in five months.
• Anthropic says Scion is "by no means a suitable replacement for git for general use" and has not announced a release.

Why git stopped fitting

The numbers in Anthropic's post are the story. The main branch history is almost 100 GB and a compressed working copy over 25 GB. The company runs roughly 25 million checkouts a day for continuous integration alone, before developer machines, agent sandboxes and research jobs. "Fetched naively, that would be over 600 PB of git traffic per day," the authors write, and at peak they were running several hundred internal git mirrors to cope.

The tenfold rise in merges is attributed to Claude writing code autonomously. The post's line is that "the latent demand for merging code is practically infinite", which is a statement about agents, not people: a model that can open a change request every few seconds will saturate any system designed around humans typing.

What Anthropic Scion does differently

Git assumes a client that works offline and holds a complete copy of the repository. Scion assumes the opposite: clients are always online, and most of them are agents and CI jobs rather than engineers. The server answers questions about history and blame, creates commits, and performs merges and rebases. The client keeps a bounded cache and a working copy that is "optional, and sparse by design". Many jobs need no checkout at all; they ask the server for a diff through an API.

The unit of work is a changeset recorded in a server-side log rather than a commit pointing at a tree, so "there's nothing to push". Data is content-addressed and spread across caching tiers, each of which the post says holds a cache hit rate of over 99%. There is no full copy of the repository anywhere, including on the server.

The reported results: an infrastructure footprint 200 times smaller than the git mirroring it replaced, checkouts for CI and research that dropped from minutes to seconds, and a 25% gain in CPU capacity for CI on its own. Those are Anthropic's own measurements of a system nobody outside can run, which is worth remembering when they are repeated as benchmarks.

How review changed

Antique, the companion review tool, can express "complex semantic constraints on approvals", so a comment-only change and a change to sensitive logic are treated differently. Engineers still review every change request, but the post says they now "examine the intent of the change as opposed to reading every line of code".

The team also dropped the idea of a global merge queue. Instead of a lock that serialises every merge, they rely on post-merge and pre-deploy testing and accept that concurrent changes sometimes conflict, noting that Claude fixes those conflicts faster than humans notice them. Not everything the model did was welcomed: the post records engineers having to tell Claude "no" when it violated system constraints or over-engineered a fix in response to an adversarial review.

One figure will travel furthest. Given a handwritten git-to-Scion ingestion prototype and "no further instructions", an early internal snapshot of Claude Mythos Preview produced a version that used 20 times less memory and ran four times faster. That is a single anecdote from the vendor about its own model, reported here as such.

What it means if you use Claude Code

Nothing changes in the tool. Claude Code still works on ordinary git repositories, and so do GitHub Copilot and Cursor. The post is a description of what happens inside a company where an agent writes most of the code, and the practical lesson for a smaller team is in the review section, not the storage layer: review intent rather than lines, invest in tests that run after merge, and expect to say no to the agent regularly.

The post ends with suggestions for git platforms rather than a product: comprehensive server-side RPCs that the git client could learn to talk to, a negotiation protocol that does not have to look like today's, and sparse and shallow features that "compose well". The authors say the concepts "really only make sense for truly enormous monorepos".

For the tools this was built around, see Claude vs GitHub Copilot, Claude vs Cursor and the ranked best AI coding tools.

FAQ

Did Anthropic stop using GitHub?

The post says Anthropic moved its monorepo off git, the version control system. It does not describe leaving GitHub as a hosting service, and its closing section is addressed to git platforms as potential adopters of Scion's ideas.

Can I use Anthropic Scion?

No. Anthropic has not released it and says it is "by no means a suitable replacement for git for general use". The post frames it as a case study for very large monorepos with agent-scale merge volume.

What is Antique?

Anthropic's in-house code review tool, built by the same seven-person team. It enforces semantic constraints on approvals so reviewers can concentrate on changes that matter, and the post says it could warrant a write-up of its own.

Does this change how Claude Code works for me?

No. Claude Code works on git repositories as before. The post describes Anthropic's internal infrastructure, not a change to the product.

How much code at Anthropic does Claude write?

The post attributes the rise from 2,500 to over 30,000 merges a day largely to Claude writing code autonomously, without giving a percentage. Earlier third-party reports have cited higher figures; this article uses only what the post itself states.