How Anthropic Replaced Git with Scion Using Claude

Claude broke Git before it replaced it

On 13 October 2026, Anthropic announced that its monorepo no longer runs on Git. The replacement is Scion, a version control system that a team of seven built in-house and moved the monorepo onto in five months.

The reason fits in two numbers. In May 2026, Anthropic averaged about 2,500 merged code changes a day. On 9 September, it merged 30,401, its record day.

And Anthropic blames Claude. Its ability to write code was, in the post's words, "what put us in this situation in the first place." Another Anthropic page says Claude authored more than 80% of the code it merged as of May 2026.

So here's my take. The story isn't that an AI wrote a version control system. It's that AI made code so cheap that the tools around the code broke. Git was just the biggest crack.

Why Git stopped fitting

Git gives every clone a full copy of the repo that works offline. That's a big part of why it won. At Anthropic's size, it's also the problem.

The main branch history alone is almost 100 GB, while CI runs on the order of 25 million checkouts a day. The team tried the standard fixes: shallow clones, sparse checkouts, internal mirrors. At the peak, they ran several hundred Git mirrors just to keep up.

Then they noticed something obvious. Nearly every client was an agent or an automated job. Those are always online. So why ship each one a full offline copy?

How Scion works

Scion turns Git around. The server answers blame and history questions, creates commits and runs merges, while the client just keeps a small cache. It fetches files only when something reads them.

Four choices carry most of the design:

• No local commits. Every commit, rebase and merge runs on the server. There's nothing to push.
• Changesets. Each one stores only the files that changed, plus pointers to their content.
• Optional checkouts. A CI job can ask the server for a diff without touching the disk.
• Caches in layers. Each cache tier keeps a hit rate over 99%.

The payoff, per Anthropic: Scion needs 200x less CPU and memory than the old Git mirror fleet. The text says checkouts got two orders of magnitude faster. Its own chart is more modest. A median CI checkout fell from 31 seconds to 3.5, about 9x, while a fresh agent worktree fell from 7.1 minutes to 5.1 seconds.

Where Claude actually fit

The best example is the importer. The team mostly hand-wrote the first Git-to-Scion ingestion pipeline, then gave it to an early internal snapshot of Claude Mythos Preview for a rewrite. With no further instructions, it came back using 20x less memory and running 4x faster.

But read the next part closely. Every engineer on the project "would now rate Claude as better at writing code than them." Writing. Not designing.

Very often, the post says, "the role of the engineers was to tell Claude 'no.'" Agents sometimes broke core system constraints. Adversarial reviews sometimes pushed Claude to over-engineer. Humans still review every change, but they check its intent instead of every line.

One catch. Odds are you can't use the model that did that rewrite, because Anthropic limits Mythos-class access to a small set of vetted organizations.

The best objection: Meta built this first

On 14 October, sunshowers.io published "Anthropic's Scion is heavily inspired by Mononoke," which reached 426 points on Hacker News. Mononoke, the server behind Meta's Sapling source control, uses "Bonsai" changesets that look a lot like Scion's.

Its author, Rain, worked on Mononoke at Meta. The charge hits this post's title head-on: Meta's team "came up with many of the insights that Claude dutifully applied almost a decade later."

Rain didn't object to the copying. "The problem is attribution, not copying." The original post named Sapling once and never mentioned Mononoke. Citing private correspondence, Rain also wrote on Hacker News that "Claude was pointed directly at Mononoke and asked to re-engineer it." Nobody outside can verify that.

Nick G., the post's author, replied on 15 October: "We also did not just tell Claude to build us a version control system." Asked how much of Scion's code was directly inspired by Mononoke, the answer was "Very little, overall." After Anthropic added a footnote crediting Mononoke that same day, Rain accepted it as enough credit.

Even the scale isn't new. A 2016 paper on Google's monorepo counts 16,000 human changes and 24,000 automated ones on a typical workday.

My read: both sides are right. Much of the design has prior art at Meta and Google. Anthropic credited both companies on day one and named Mononoke after the backlash. The speed was Claude's. Claude moved fast because seven people knew which proven design to aim it at and when to say no.

Your Git is fine

Anthropic says it plainly: "Scion is by no means a suitable replacement for git for general use."

Here's my own receipt. Awiser's two repos hold 1,273 commits on their main branches in total, with 30 on the busiest day on record. At its record pace, Anthropic merges that much in about an hour.

Git isn't my bottleneck. It's almost certainly not yours.

The lesson under the chart still applies: when code gets cheap, the bottleneck moves to whatever comes after it. For Anthropic, that was source control, builds and CI. For a solo builder, I think it's users. Claude can write the feature. It can't find the people who need it.

How Awiser fits

Disclosure: I build Awiser, so I'm biased toward "the bottleneck is users." Its growth copilot gives you one move at a time, matched to your project's stage. At stage one, the move is talking to real users, not launching. Past that, here's where I'd launch a side project.

One thing to do tonight

Run git rev-list --count --since=30.days.ago HEAD in your main repo. Divide by 30. That's your commits per day.

Anthropic's record is 30,401. If yours is under a hundred, close the Scion tab. Spend the hour on one user instead, then ask the copilot for your next move.