Skip to content
Mimir Lab
← Back to blog

how it works

Did Anthropic replace git? Only in its own monorepo: what Scion is, and what it isn't

· anthropic, claude-code, git, source-control, ai-agents, ci

In short: on 13 October 2026 Anthropic published Redesigning source control at Anthropic. It built Scion, an in-house version control system, and “fully migrated our monorepo from git in five months”. That is the whole replacement: one repository, Anthropic’s own. The same post says Scion “is by no means a suitable replacement for git for general use”. There is no public code, and Claude Code still runs on git.

Within hours, “our monorepo” had dropped out of the headlines. Below: what was announced, why git stopped fitting, how Scion works, what it isn’t, and what a small team can reuse without leaving git. Facts first, with sources; my read in a section of its own.

Research: a three-agent Claude Code run across the primary post, Claude Code’s changelog and docs, Anthropic’s public repositories and two days of coverage. Every quote below was checked against the original page.

What Anthropic announced

The post is on Anthropic’s engineering blog, dated 13 October 2026 and signed by Nick G. and the source control team (Anthropic). A team of seven built Scion, a code review tool called Antique, code search, a source control IAM system and “half a dozen related systems”, and migrated to them in five months.

The trigger is a number: “In May 2026, Anthropic averaged around 2,500 code changes merged per day; as of September, we merged over 30,000 per day.” The post calls source control “our biggest development bottleneck” and doesn’t expect the growth to slow down.

Claude Code isn’t mentioned. Two internal models and tools are: Claude Mythos Preview, which rewrote the git-to-Scion ingestion pipeline into a version that “consumed 20x less memory and was 4x faster”, and Claude Tag, which the team uses to watch each component of the system.

Why git stopped fitting

Anthropic’s description of git is respectful. Every clone is a complete repository that works offline, and “for the vast majority of repositories, this is the right design”. Their monorepo left that design space:

Metric Value (Anthropic’s figures)
Main branch history almost 100 GB
Compressed working copy over 25 GB
CI checkouts on the order of 25 million a day
Git traffic if fetched naively over 600 PB a day
Internal git mirrors at peak several hundred

They had already used the standard workarounds: shallow clones, sparse checkouts, promisor remotes, mirrors, repository snapshots in blob storage. These “bought us time”, then got harder to operate as the merge rate climbed.

The key observation is in one sentence. Like most companies, they used git as a centralised system with one authoritative copy, but “nearly all of our clients were agents and jobs that are online by definition”. The properties that make git good offline had become the cost.

How Scion differs from git

All from the post unless marked.

git Scion
Client Full repository, works offline Always online, bounded local cache, computes very little on its own
Commits, merges, rebases Local, then push On the server, nothing to push
Unit of work Commit pointing to a tree “Changeset”: a write-ahead log entry listing the files changed against its first parent, with content-addressed pointers into blob storage
Refs (branches, tags) Everywhere Discouraged except main and release branches; moving one is “a compare-and-swap on a single Spanner row”
Storage Packfiles, periodic maintenance Distributed, no full copy anywhere, no packfiles, no repository maintenance
History and blame Computed locally Server APIs over “derived data” built from changesets
Fetch Negotiates what the client already has Client asks only for the trees and blobs it wants, so sparse and virtual checkouts need no special support
Merge queue A global lock is the industry standard No global lock; post-merge and pre-deploy testing instead

Checkouts go through three tiers that answer identical RPCs: a daemon on each node, regional caches, and the authoritative servers. Each tier reports a cache hit rate above 99%.

The reported results: an infrastructure footprint “200x smaller” in CPU and memory than the old git mirrors, checkouts “two orders of magnitude faster (in seconds rather than minutes)”, and a 25% CPU capacity gain in CI. These are Anthropic’s numbers. I found no external measurement.

On conflicts, Nick G. clarified on Hacker News: “We never allow textual conflicts”, while semantic conflicts are “an unavoidable fact of very large repos” unless merging is fully serialised.

Anthropic names its prior art: Google’s Piper and Meta’s Sapling. Not everyone found that enough. Rain, who helped build Meta’s Mononoke source control server, wrote on sunshowers.io that Scion’s “derived data” concept mirrors Mononoke’s, and that crediting Sapling only for “hard-won lessons” while not mentioning Mononoke “borders on plagiarism”.

What Scion is not

  • Not released. I found no Scion repository in Anthropic’s GitHub organisation and nothing on npm. The post explains why: “it’s difficult to offer Scion in a form that would be broadly useful to other organizations”, because it “requires its own code review, code search, and CI”.
  • Not part of Claude Code. Claude Code’s changelog, up to version 2.1.300, doesn’t contain the word “Scion”. Worktrees still “require a git repository” (docs).
  • Not Claude Code’s checkpoints. /rewind and checkpoints have existed since Claude Code 2.0 in September 2025 and are a per-session undo buffer. The docs give it its own heading: “Not a replacement for version control” (docs).
  • Not Anthropic leaving git everywhere. The post talks about “our monorepo”. Anthropic’s public repositories are still git repositories on GitHub.
  • Not the other Scion. Google has an unrelated open-source agent orchestration project with the same name (GoogleCloudPlatform/scion). Third-party “open Scion” reimplementations of the published design appeared on GitHub within a day; none of them is Anthropic code.

How “our monorepo” disappeared

The post reached Hacker News at 15:49 UTC on 13 October under its own title and collected 879 points and nearly 1,000 comments (HN). Careful coverage kept the scope. The Register headlined “Anthropic reinvents source control as Claude floods the world with code” and opened with git’s approach being “the right one for most code repos, according to Anthropic” (The Register). GitLab wrote that Anthropic “acknowledges that its approach isn’t a general-purpose replacement for Git” (GitLab).

On social media and aggregators the qualifier went first. aiweekly.co is a clean example: its body correctly says Anthropic “has replaced Git in its monorepo” and that Scion will not be open-sourced, while the same page lists an “Original headline” of “Anthropic Open-Sources Scion, a Git Replacement Built by Seven Engineers With Claude in Five Months” (aiweekly). Two words dropped, one claim added, and the story flips.

The post’s author answered on HN the next day: “The point explicitly isn’t that you should vibe-code your own VCS. There are a lot of ideas here that would be directly applicable to git” (HN).

What a small team can reuse without leaving git

Anthropic’s own closing points here: apply “some of the most successful parts of Scion to git without departing from the overall model”. At small scale that comes down to three habits.

Count checkouts, not commits. Every CI job, agent sandbox and parallel session is a clone. Anthropic’s pressure came from 25 million checkouts a day, not from the commits themselves. When agents run in parallel, the number to watch is how many times a day the repository gets pulled, and how much of it each job actually reads.

Fetch only what the job reads. Git already has the tools; they just aren’t the default everywhere.

# commits and trees only: file contents download on demand,
# and only the files at the repository root are checked out
git clone --filter=blob:none --sparse https://github.com/OWNER/REPO.git
cd REPO

# add just the folders this job touches
git sparse-checkout set src/content docs

In GitHub Actions, actions/checkout already defaults to fetch-depth: 1 and accepts a sparse-checkout list. This site’s own CI deliberately uses fetch-depth: 0 because gitleaks scans the full history. That is right at this size, and worth revisiting the day the repository stops being small.

Ask the server instead of cloning. Scion exposes its queries as APIs, so many jobs need no local repository at all. GitHub already covers part of this: the compare endpoint returns the diff between two refs without a clone.

curl -H "Authorization: Bearer $GITHUB_TOKEN" \
  https://api.github.com/repos/OWNER/REPO/compare/main...feature-branch

In n8n that is a single HTTP Request node. For an agent that only needs to review a diff, it replaces a checkout.

My read

From here on it’s my opinion, kept apart from the facts above.

The headline is wrong; the trend under it isn’t. When writing code gets cheap, the bottleneck moves to everything after it: merging, CI, review. Anthropic hit it first because, in its own words, its “ability to spend tokens for internal use far exceeds what any other business would realistically want to spend”. It expects “this class of problems” to become common within six to twelve months.

I don’t expect a new VCS to install. I expect the change to land on the server side of git. GitLab announced a next-gen SCM for agent workloads in June, with Nick G. from Anthropic’s source control team on stage, and is keeping git compatibility (GitLab). Anthropic writes that it wants to “partner with GitHub, GitLab, and the OSS community”. My bet: within twelve months the large forges ship server-side query APIs and leaner fetch aimed at agents, and the git client on my laptop stays the same.

The line I’d keep from the post isn’t technical: “Very often, we found that the role of the engineers was to tell Claude ‘no.’” Their reviewers now “examine the intent of the change as opposed to reading every line of code”. That applies to a small team as much as to Anthropic.

For a Swiss SMB with a handful of developers and a few agents, there is nothing to do on source control. Measure CI minutes and checkout time the week you start running agents in parallel: that’s where it shows first.

What we still don’t know

As of 15 October 2026:

  • Whether Scion will ever be released or licensed. The post calls it hard to offer; no plan is announced.
  • Whether Scion can mirror to or from git remotes now that the migration is done. During the migration, Anthropic “ran the repo ingestion in a loop for several months” to keep the two in sync (HN).
  • Whether Anthropic’s other internal repositories are still on git.
  • How Antique, the code review tool, works. The post says it could fill “a whole other blog post”.
  • Any external measurement of the 200x, seconds-not-minutes and 25% figures.

Sources

https://www.anthropic.com/engineering/redesigning-source-control

What it means