Anthropic Scion: the in-house source control system that replaced Git for its monorepo (and why it isn't open source)

On October 13, 2026, Anthropic published an engineering post describing Scion, a version control system it built to replace Git for its main code repository. The reason is volume. Anthropic says it merged about 2,500 code changes a day in May 2026 and more than 30,000 a day by September, as Claude did more of the coding. Scion isn't open source, though. The post contains no code, repository or license, and Anthropic says Scion isn't meant to replace Git for anyone else.

This story is based on Anthropic's engineering post, "Redesigning source control at Anthropic," checked October 14, 2026. Every number below is Anthropic's own.

The short version

• What it is: an internal, server-centric version control system for Anthropic's monorepo, with a client called scn. It also comes with a new code review tool called Antique.
• Why: "As our merges per day grew more than tenfold, source control became our biggest development bottleneck," Anthropic writes.
• Claimed results: infrastructure 200x smaller in CPU and memory than the old Git mirrors; checkouts in seconds instead of minutes; a 25% CPU capacity gain in CI.
• Open source? No. There's no repo or license, and Anthropic says it's "difficult to offer Scion in a form that would be broadly useful to other organizations."
• Can you try it? No. Anthropic is sharing design ideas it hopes Git platforms will borrow.

Why did Anthropic replace Git?

Git was designed for developers who work offline with a full copy of a repository. Anthropic says that assumption no longer fits its traffic. Nearly all of its "clients" are now always-online agents and CI jobs, not people at laptops.

The post gives the scale:

Last verified: October 14, 2026 (all figures from Anthropic's October 13 post)

To keep up, Anthropic ran "several hundred internal git mirrors" at peak. Those mirrors got harder to operate as the merge rate climbed. Anthropic ties the growth directly to its own models: "Claude's ability to write code was both what put us in this situation in the first place" and what let a small team fix it. If you've wondered what agents running unattended at scale do to the rest of the toolchain, this is one answer. Our piece on the case against unattended agents covers the other side of that trade-off.

How is Scion different from Git?

Scion moves the work from the client to the server. According to the post:

• Changesets, not packfiles. The basic unit is a "changeset" written to a log on the server. File contents live in blob storage. History, blame and commit-graph indexes are derived from those changesets. With no packfiles, there's no repository maintenance to run.
• The server does the heavy lifting. Clients are assumed to be online. The server handles history, blame, commits, merges and rebases. The client keeps a limited local cache and lays out the working copy.
• Files on demand. Scion supports sparse checkouts and "virtualized checkouts, which materialize files on demand," so a job only pays for the files it actually reads.
• A cache hierarchy. The pieces are the scn client, a per-machine scion-daemon, regional scion-cache servers, and authoritative scion-servers, with data in Spanner and cloud storage. Anthropic says each cache tier keeps a hit rate above 99%.
• Fewer branches. Anthropic's workflow discourages refs other than main and release branches, because global state that every server must agree on is "a looming scaling problem."

Anthropic also built Antique, a code review tool that can express "complex semantic constraints on approvals" and runs the merge queue. The post says a team of seven built both Scion and Antique, and that the full migration took five months.

What does Anthropic say it gained?

• Infrastructure: "Our Scion infrastructure footprint is now 200x smaller than our previous git mirroring infrastructure in terms of CPU and memory requirements." That compares server resources against the old mirrors. It isn't a speed figure.
• Checkouts: "CI and research jobs check out code two orders of magnitude faster (in seconds rather than minutes)."
• CI capacity: "CI alone realized a 25% CPU capacity win."
• An AI rewrite: Anthropic says Claude rewrote the pipeline that brings Git history into Scion, and the new version used 20x less memory and ran 4x faster than the hand-written one.

None of these numbers has been independently checked. They describe one company's monorepo and infrastructure, so they don't predict what any other team would see.

What are the trade-offs?

The post is candid about costs:

• No global merge lock. Anthropic says it hasn't been able "to effectively replicate the industry standard of a global repository lock on the merge queue." It relies on testing after merge and before deploy instead, so changes that land at the same time can conflict.
• Everything else had to change too. Because Scion diverges from Git, adopting it meant replacing code review, code search and CI integration.
• AI help had limits. Engineers set the design and told Claude when its proposals broke constraints. The post calls AI code review "myopic" at times. It also notes Anthropic's internal token spend "far exceeds what any other business would realistically want to spend."

Is Scion open source, and can you use it?

No. The post doesn't mention a code release, a repository or a license. Anthropic says Scion diverges so far from Git that adopting it would mean replacing code review, code search and CI too, and that its internal versions of those components "are highly bespoke, so they would be difficult for other organizations to adopt." It adds that "Scion is by no means a suitable replacement for git for general use." The ideas, it says, "really only make sense for truly enormous monorepos."

What Anthropic offers instead is a direction for existing tools. It hopes developers will "apply some of the most successful parts of Scion to git," notes that "a git server doesn't have to work the same way as a git client does," and says it wants to "partner with GitHub, GitLab, and the OSS community." It names Google's Piper and Meta's Sapling as prior art.

One naming warning: a GitHub project called Scion under the GoogleCloudPlatform organization is unrelated. It's an Apache 2.0 tool for running coding agents such as Claude Code and Gemini CLI in isolated containers, and its page says it isn't an officially supported Google product.

What it means for you

For almost every team, nothing changes. Git and the major forges handle repositories far smaller than Anthropic's. The useful signal is directional. If coding agents keep multiplying the number of changes per engineer, the bottleneck moves to merging, review and CI. Our comparison of Claude Code, Codex and Cursor covers the tools producing that volume, and our Anthropic explainer covers the company behind it.

What's confirmed, reported and uncertain

• Confirmed (Anthropic's post, October 13, 2026): Scion exists and replaced Git for Anthropic's monorepo; the merge, repository-size and checkout figures; the architecture; the 200x footprint, checkout-speed and CI-capacity claims; the seven-person team and five-month migration; that Scion isn't offered publicly.
• Not supported by any primary source: claims that Anthropic open-sourced Scion. Anthropic's post announces no repository, license or release.
• Uncertain: whether Anthropic will ever publish code from Scion or Antique, and whether GitHub or GitLab will adopt any of its ideas. Anthropic also doesn't say whether its engineers still use Git for its public open-source projects.