Anthropic replaces Git with in-house Scion as Claude-generated code overwhelms repositories

AI coding volume pushes Anthropic to build its own monorepo source control system

By LineZotpaper
Published
Read Time3 min
Anthropic has built a homegrown version control system called Scion after Claude models generated so much code that the company's existing infrastructure could not keep up. The company says it was merging more than 30,000 code changes per day as of September, up from about 2,500 per day in May, and it expects that growth to continue.

Anthropic detailed the move in a blog post on Tuesday, describing a five-month migration to Scion, a centralized monorepo system that in some ways resembles Subversion rather than Git's distributed architecture. The company said its main branch is about 100 GB, with a compressed working copy more than a quarter of that size, and that it performs about 25 million code checkouts per day for continuous integration alone. Fetched naively, it said, that would mean more than 600 PB of Git traffic per day.

"In May 2026, Anthropic averaged around 2,500 code changes merged per day; as of September, we merged over 30,000 per day," the company said. "We don't expect this rate of growth to slow down anytime soon; the latent demand for merging code is practically infinite, and every bottleneck we remove immediately leads to a further increase in the amount of code we produce."

Anthropic is not the first organisation to hit the limits of Git. Google uses its own Piper monorepo system, Microsoft built Virtual File System for Git, and Meta developed Sapling. GitHub, which is owned by Microsoft, said in April that recent availability problems were linked to the exponential growth in pull requests and commits. Academic researchers have also noted the flood of AI-generated code.

According to Anthropic, Scion was designed with AI agents in mind. The client leaves most of the work to the server, source control queries are exposed through an API, and the unit of work is a "changeset" stored in a write-ahead log rather than a Git commit pointing to a tree. Claude also helped build related tools, including Antique, a code review interface, a code search tool, and a source control IAM system. Work on Scion began in January, and when the Claude Mythos Preview became available internally, the model rewrote an early Git-to-Scion ingestion pipeline that used 20 times less memory and ran four times faster.

Anthropic acknowledged that the decentralized approach embodied by Git remains the right one for most code repositories. But at the company's scale, engineers found it impractical. Developers who worked on Scion, the company said, agree that Claude writes better code than they do, but they keep a role as reviewers: "Very often, we found that the role of the engineers was to tell Claude 'no,'"

§

Analysis

Why This Matters

  • AI coding tools are shifting the bottleneck from writing code to integrating and reviewing it, with implications for how software teams are organised.
  • The scale cited by Anthropic, 30,000 merges per day, points to a future where source control and continuous integration systems must serve machine agents as much as humans.
  • Other large AI-driven organisations may face similar pressure to adapt or replace Git-based workflows.

Background

Git became the dominant version control system for its distributed, branching-friendly design, but large companies have long pushed its limits. Google runs a monorepo on internal systems, and Microsoft and Meta built scaling tools around Git. Centralized version control such as Subversion was the norm before Git. Anthropic's move is notable because the company says the pressure is coming from AI agents, not just a growing number of human developers.

Key Perspectives

Anthropic: Treats AI-generated code volume as effectively unbounded, and sees centralized, server-side source control as the practical way to manage it. Engineers still matter as reviewers who can say no to Claude's suggestions. Traditional open-source community: Git's distributed model has been widely adopted and remains appropriate for most projects, as Anthropic itself acknowledges. The company's solution is tailored to an extreme scale that few organisations will reach. Critics and skeptics: The rapid growth in merges raises questions about review quality and the operational cost of running CI at this scale. There is also a risk in using AI to build the infrastructure that manages AI-generated code, if errors compound rather than are caught.

What to Watch

  • Whether Anthropic's merge rate continues to climb after the Scion migration and how its CI infrastructure copes.
  • Whether Scion remains internal or is eventually released or documented for other organisations facing similar loads.
  • How Git-based tooling, including GitHub, responds to agent-driven development and whether it adopts server-side or API-first features.

Sources

Zotpaper

Written by software from the reporting listed above, scored by an automated standards desk, and published without a person reading it first. If something here is wrong, tell the editor and it will be put right.

How we workSubscribe