Anthropic replaces Git with in-house Scion source control as daily merges pass 30,000

• Scion: internal source control for a single monorepo. Commits, merges and history queries run on servers. Clients fetch files on demand into a bounded local cache, and jobs can query the repo through APIs without keeping a local copy (Anthropic Engineering).
• Antique: internal code review. Approval rules can be semantic, for example telling comment-only edits apart from sensitive logic changes, so human reviewers focus on the riskier changes.
• Models: the post names Claude Mythos Preview and "Tag" as having driven the need for Scion and helped build it.
• Access: none. Neither system is open-source or available as a product.

• Scale: main-branch history is about 100 GB, and a compressed working copy is over 25 GB. CI runs about 25 million checkouts a day, which would mean more than 600 PB of Git traffic a day if each one fetched naively (Anthropic Engineering).
• Data model: "changesets" replace Git commits. Each is an entry in a write-ahead log that stores only the changed files plus pointers to content-addressed file data. There are no packfiles, no locks and no coordination between servers. Derived data such as blame and path history is computed and cached separately.
• Serving: three cache tiers: a local daemon, a regional cache, then the authoritative servers. All tiers answer the same requests, and Anthropic reports cache hit rates above 99%. No server holds a full copy of the repository.
• Merging: there is no global repository lock. Anthropic relies on testing after merge and before deploy, with Claude resolving conflicts automatically.
• Reproducibility: no code, benchmarks or test setup have been published.

• Infrastructure: Scion uses 200 times less CPU and memory than the previous Git mirroring setup (Anthropic Engineering).
• Checkouts: about 100 times faster, taking seconds instead of minutes, which saved about 25% of CI CPU.
• Ingestion pipeline: Claude rewrote the Git-to-Scion pipeline to use 20 times less memory and run 4 times faster.

All of these figures are reported by Anthropic only. Secondary coverage repeats them without new measurements (Crypto Briefing; AlphaSignal). Simon Willison links the rise in velocity to Fable/Mythos-class models, but he frames that as his own guess (Simon Willison).

Anthropic says it learned from Google's Piper and Meta's Sapling. Piper is a centralized system built on Bigtable and later Spanner. In 2016 it held about 86 TB and 35 million commits, and engineers used it through cloud workspaces that store only modified files (CACM). Sapling began as an extension to Mercurial and was released as open source in 2022 as a client that works with Git (Meta Engineering). The difference with Scion is the main user: agents rather than human engineers.

• No offline mode. On Hacker News, a Scion engineer said local commit buffering would mean copying server logic onto the client. It was not a priority because Claude, by far the heaviest user, is always online (HN comment).
• Over-engineering. Adversarial review by Claude tended to over-engineer in response to hypothetical concerns, so engineers had to rein it in (Anthropic Engineering).
• Skeptical commenters. Some on Hacker News questioned the 30,000-a-day and 100 GB figures. Others pointed to visible quality problems in Anthropic's products as a sign of technical debt at this pace (Hacker News).

• Watch the bottleneck move. As agent output grows, measure capacity in CI, review queues and merge workflows, not just how fast code gets written (DEV Community).
• Most teams shouldn't copy this. Anthropic says Scion only suits huge monorepos and needs tightly integrated review, search and CI. Its authors would rather see Git hosting platforms add server-side APIs and better sparse and shallow checkouts (Anthropic Engineering).
• The engineer's role changes. At Anthropic, engineers increasingly checked intent and design and turned down Claude's proposals, rather than reviewing code line by line.