Skip to main content
Back to timeline
AnthropicSource publication:

Anthropic moves its monorepo off git onto in-house Scion: merges rise tenfold while infrastructure footprint shrinks 200-fold

Synopsis

Anthropic has described Scion, its in-house version control system, and completed a full migration of its monorepo off git in five months. The trigger was AI agents writing code: daily merges rose from about 2,500 in May 2026 to over 30,000 in September, while git's assumption of complete offline copies and several hundred internal mirrors struggled with roughly 25 million code checkouts a day. Scion keeps clients always online, moves commits and merges to the server, and replaces full working copies with sparse and on-demand virtualized checkouts; its infrastructure uses 200 times less CPU and memory than the previous git mirroring setup, and CI checkouts drop from minutes to seconds.

Source-provided article image: Redesigning source control at Anthropic \ Anthropic
Anthropic

Interpretation

Anthropic replaced git with its own Scion for version control of its monorepo, shipping alongside it a code review surface called Antique, code search and a source control permissions system, and completing the full migration in five months. Anthropic had run the monorepo on git and relied on shallow clones, sparse checkouts, promisor remotes, internal git mirrors and repository snapshots in blob storage, operating several hundred internal git mirrors at peak. Main-branch history alone is almost 100 GB and a compressed working copy exceeds 25 GB, with about 25 million code checkouts per day that would exceed 600 PB of git traffic daily if fetched naively; daily merges rose from about 2,500 in May 2026 to over 30,000 in September.

Scion treats clients as always online with only a bounded local object cache, and performs commits, rebases and merges on the server, so there is nothing to push. git assumes every clone is a complete repository that works offline and every client can act as a full server, and those properties designed for distributed offline work limited scaling as the repository grew. The unit of work is a changeset in a write-ahead log that records files changed relative to its first parent plus content-addressed pointers to compressed contents in blob storage; blame, path history and commit-graph skip lists are derived data computed by one recursive mechanism and can be deferred.

Checkouts fetch objects on demand: the client starts from a root tree and recursively requests missing trees and blobs, and each request targets an immutable object any cache tier can answer. git negotiation tells the server what the client already has and can return a single multi-gigabyte packfile response, whereas Scion negotiation operates only on the trees and blobs the client wants. A node-local scion-daemon starts from a periodically refreshed snapshot and talks only to regional scion-cache servers, which fall back to authoritative scion-servers on a miss; each tier keeps a cache hit rate above 99% and needs no long-lived state.

Perspective

The system targets Anthropic's own monorepo scale and agent-dominated clients; the text states that Scion is by no means a suitable replacement for git in general use, and that its internal review, search and CI components are highly bespoke and hard for other organizations to adopt.

Anthropic says it has not been able to replicate the industry-standard global repository lock on the merge queue, and instead invests in post-merge and pre-deploy automated testing, so concurrent changes can still conflict; it also notes that its solution and considerations may not suit every case.