Anthropic's Scion copies Mononoke, says one of its builders

Ask Opus 5.5 or GPT-6 Astra about prior art for Scion, Anthropic's new version control system, and both models name Mononoke right away. That's the account in a post on Sunshowers by an engineer who helped build Mononoke, the source control server developed at Meta, who says Anthropic's write-up on Scion "borders on plagiarism".

At a glance

• Anthropic's engineering post on Scion uses "derived data" to make blame faster. The author described the same idea for Mononoke in 2022 Oxide candidate materials, based on a Meta team design from 2017-18.
• Derived data means a result like blame is computed once and stored. Later requests are answered from the saved result, so the server doesn't redo the work every time.
• The author is fine with the copying itself, since Mononoke is open source. The objection is to the credits: Sapling got a mention only for "hard-won lessons", and Mononoke got none.

In case you missed it, Anthropic recently published a post explaining why it replaced git internally. According to Anthropic, writing code is no longer the bottleneck. The bottleneck is developer infrastructure built for human scale. The company says it averaged around 2,500 merged code changes per day in May 2026 and more than 30,000 per day as of September. Before Scion, it tried git workarounds and at its peak ran several hundred internal git mirrors.

Anthropic and the Mononoke author both use "derived data" to explain fast blame

The author's evidence is two excerpts. One is Anthropic's description of how Scion uses derived data to make operations like blame more efficient. The other is the author's own 2022 description, written for Oxide candidate materials, of how Mononoke uses the same concept for the same job.

The author says the imitation goes well beyond similar insights. Anthropic's post credits Sapling only for "hard-won lessons" and doesn't mention Mononoke at all. As the author reads it, the post comes off as Anthropic claiming that Claude designed everything. In the author's view, the Meta team came up with many of the insights that Claude applied almost a decade later.

The author also points out that the Meta team had prior art of its own, from Google's SrcFS to write-ahead logs in databases. Their wider argument: as LLMs help design new systems, they will lean heavily on existing designs, and Mononoke is open source and sits in every model's training set. That makes giving credit easier than before, and more important.

Seven engineers built Scion and moved Anthropic onto it in about five months

A team of seven engineers built Scion and migrated to it in about five months. In the same period they built Antique, a code review surface, plus code search and source-control IAM. Anthropic hasn't announced a public release plan. It presents Scion as a case study and says it intends to partner with GitHub, GitLab and the open source community.

Anthropic gives these scale figures: almost 100 GB of main-branch history, a compressed working copy of more than 25 GB, and on the order of 25 million CI checkouts per day. Fetched naively, that would come to more than 600 PB of git traffic per day.

According to Explainx, the chart in Anthropic's post shows a single-day peak of 30,401 merges on 2026-09-09. Explainx also lists the company's reported results: a CPU and memory footprint 200x smaller than the old git mirror fleet, checkouts two orders of magnitude faster, and a 25% CPU capacity win in CI. Anthropic reported these figures itself, and nobody has verified them independently.

In git every clone carries the full history, and Anthropic says its clients are always online

In git, every clone is a complete repository. It can compute diffs, read history and blame, and check out any revision offline. Anthropic argued that nearly all its clients, mostly agents and jobs, are always online, so this design worked against scaling.

According to AlphaSignal's summary of Anthropic's account, Scion draws on ideas from Google's Piper and Meta's Sapling. Clients keep only thin state and fetch data on demand. Commits, merges, rebases, diffs, ancestry queries and blame all run on the server. Storage is distributed, with no packfiles and with sparse virtualized checkouts.

NexusAI calls derived data the mechanism behind fast server-side blame. In its description, Scion builds blame indices, path histories and skip lists from changesets and stores them incrementally. Each can be computed eagerly or lazily, and that choice affects latency, storage and cost. It works like a book index: someone compiles it once, and every reader can find a word without scanning every page.

The evidence published so far covers one concept, shown as two excerpts. The author doesn't compare the wider architecture point by point, and even cites Vicent Marti's post for Cursor as a system with similar insights but many key design differences. In our view, the missing credit is odd in a post about Claude doing design work, because by the author's account a single question to a model turns up Mononoke.

Whether Scion ever leaves Anthropic Anthropic hasn't given a public release date for Scion, only its plan to partner with GitHub, GitLab and the open source community. If that work carries Scion's ideas into outside tools, the question of credit will follow them. It isn't known whether Anthropic will revise its post to mention Mononoke. The discussion continues on Hacker News and Lobsters.