Anthropic's AI coding tools started merging so much code that git, the version control system nearly every developer uses, couldn't keep up. So a seven-person team built a replacement and moved the company's entire main codebase onto it in five months.
Why Anthropic Replaced Git With Scion
Anthropic published an engineering post on October 14 explaining why it built its own version control system, called Scion, and moved its monorepo off git. A monorepo is a single repository that holds a company's entire codebase, and Anthropic's had outgrown the design of git.
The numbers explain the squeeze. In May 2026, the company merged around 2,500 code changes per day. By September, that had climbed past 30,000. Most of that code is written by Claude, Anthropic's AI model, which the company uses to write software far more than almost any other business.
Git was built for a world where developers work offline and keep a full copy of a repository on their machines. Anthropic's history alone is almost 100 GB, and the compressed working copy runs over 25 GB. The company performs around 25 million code checkouts a day just for its automated testing systems, before counting developer setups and agent sandboxes. Handling that through git would mean over 600 PB of traffic a day, if fetched the normal way.
How Scion Redesigns Version Control for AI Agents
Anthropic's key insight was that nearly all of its clients are AI agents and automated jobs, which are always online by definition. Git's strengths come from serving offline machines, and those same traits were dragging the system down.
Scion flips the design. Instead of giving every client a full copy, it runs a daemon on each machine that stores shared objects for the whole node and talks to regional cache servers, which read directly from memory. The team says each layer of the system holds a cache hit rate above 99%, and that the infrastructure footprint is now 200 times smaller than its old git mirroring setup in CPU and memory terms.
The payoff is speed. Checks out that used to take minutes now happen in seconds, and the automated testing pipeline alone saw a 25% CPU capacity win. The company also built a code review tool called Antique, plus code search and an access control system, all in the same five-month window.
What Scion's Scale Means for AI-Written Code
The project is a case study in how much the job of engineering is shifting, and Anthropic is candid about it. All of the engineers on the team said they'd now rate Claude as better than themselves at writing code. Their value came from something else: knowing how the systems fit together and where to push back.
The team described telling Claude "no" as a recurring task. Agents, they wrote, will sometimes bend a core rule to satisfy a pattern that shouldn't exist. During the migration, several systems tried to look up when a change merged by scanning the whole commit log instead of asking the review tool. Skilled engineers steered the model away from those shortcuts and from over-engineering simple problems.
They also turned Claude loose on quality. Scion ended up among the best-tested codebases the team has worked on, and the model found optimizations, like a missing index on a query, that would've taken an expert days to spot.
Is Scion a Git Replacement for Everyone
Short answer: no, and Anthropic says so plainly. Scion "is by no means a suitable replacement for git for general use," the post states. It only makes sense for truly enormous monorepos with tens of thousands of engineers or swarms of agents writing code, and Anthropic notes its willingness to spend tokens internally far exceeds what any other business would want.
The company also won't open-source it. Scion diverges so much from git that adopting it would mean migrating your whole development lifecycle, and it needs its own review, search, and testing tools, which are highly customized to Anthropic's setup.
Where Scion fits
- Huge monorepos with agent-scale code volume
- Teams with effectively unlimited token budgets
- Companies willing to rebuild review, search, and CI around a new system
Where git still wins
- The vast majority of repositories and teams
- Anyone who needs offline access and a distributed model
- Projects that depend on the existing git tool and hosting services
What Anthropic's Scion Post Means for Developers
Even if you never touch Scion, the post flags something worth watching. Anthropic expects this class of problem to spread over the next six to 12 months as more teams lean on AI coding tools. Rather than push its own system, the company says the more useful move is to feed its ideas back to git, GitHub, GitLab, and the open-source community.
The team points to areas where git could borrow Scion's approach: server architectures that scale better, richer server-side protocols, and improved handling of sparse and shallow checkouts.
For now, Scion stays inside Anthropic. The value for everyone else is the roadmap it sketches for what version control has to become once AI agents, not people, are the main ones writing and merging code.
