Skip to content

Anthropic

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

3 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

Image: sunshowers.io

An engineer who helped build Mononoke, Meta’s source control server, says Anthropic’s October 13 write-up on Scion, its new version control system, follows Mononoke’s design so closely that the lack of credit “borders on plagiarism”.

In a post on Sunshowers, the author puts two passages side by side. In one, Anthropic explains how Scion uses “derived data” to make operations like blame more efficient. In the other, the author explains how Mononoke uses a concept with the same name for the same job, in Oxide candidate materials written in 2022. Derived data here means results like blame that the server computes and stores, so it doesn’t have to recompute them on every request. “I think the level of imitation goes well beyond having similar insights,” the author writes.

The author’s complaint is about credit. Anthropic’s post credited Sapling, Meta’s open-source source control system, only for “hard-won lessons”, and Mononoke not at all, the author says. They add that the post reads as if Claude designed it all, when “our (extraordinary) team at Meta came up with many of the insights that Claude dutifully applied almost a decade later.” They don’t mind Anthropic reusing the design: putting Mononoke out as open source was meant “to benefit humanity through the spread of good ideas.” The Meta team looked at plenty of prior art too, the author notes, from Google’s SrcFS to write-ahead logs in databases.

Mononoke’s side of the record:

  • 2016: Facebook’s Source Control team starts Mononoke as a rewrite, because the commit rate in the company’s monorepo was getting close to the most Mercurial could handle, per Meta’s engineering blog.
  • 2017: Sapling’s documentation lists “New server infrastructure for push throughput and faster indexes (Mononoke, 2017)” among its changes for large monorepos.
  • 2017-18: the author’s team works out the vision behind Mononoke’s design.
  • 2022: the author describes Mononoke’s derived data in the Oxide materials.
  • October 13, 2026: Anthropic publishes its description of Scion, derived data included.

Mononoke was built for scale. Sapling’s docs say that at monorepo scale, tens of millions of files and commits, operations that cost time or space in proportion to the number of files or commits gradually stop being affordable, so the server pre-builds various kinds of indexes ahead of time. Anthropic’s own pace in October was over 30,000 merged changes a day, and its Scion design is already being rebuilt in the open as OpenScion.

The author’s wider point is about how systems now get designed. As LLMs help design new systems, they are likely to lean heavily on existing designs, and Mononoke is open source and in every model’s training set. Ask Opus 5.5 or GPT-6 Astra about prior art, the author says (crediting Matthew Fowle), and both immediately point out the similarities to Mononoke. So the missing credit was a single prompt away.

For contrast, the author points to Vicent Marti’s blog post for Cursor, describing a system with similar insights but many key design differences.

“Giving credit to those that came before us is now easier — and more important — than ever,” the author writes. “There is no longer an excuse.”

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.