Anthropic says Git was starting to break under the weight of code produced with help from Claude. The company reported that the number of code changes merged into its main repository each day rose from about 2,500 in May to more than 30,000 in September, a roughly 12x jump in four months.
That growth led Anthropic to build its own version control system, Scion. A team of seven spent five months on the project and moved the company’s single monorepo off Git. Anthropic laid out the effort in an engineering blog post published on Oct. 13, but said Scion will not be open-sourced and will not be offered to outside users. The company wrote that it is “not a fit as a general-purpose replacement for Git” and said it shared the design mainly so GitHub, GitLab, and the open-source community could borrow workable ideas for Git’s server side.
Why Git stopped fitting Anthropic’s workload
Anthropic said the history of the main branch alone in its monorepo is close to 100 GB. Its continuous integration, or CI, systems perform about 25 million checkouts a day. If every checkout used Git’s usual full-fetch model, the company said daily traffic would exceed 600 PB.
The team tried a series of workarounds before building something new. Those included shallow clones, sparse checkouts, internal Git mirror servers, and cloud snapshots. At peak load, Anthropic said it was running hundreds of mirror servers to spread traffic, and maintaining that setup was getting harder over time.
Its core observation was that Git is built as a distributed system, with each clone holding a full copy of the repository and supporting offline work. But Anthropic said that, like most companies, it actually uses Git as if there were one central authoritative version. At the same time, nearly all of its clients are now AI agents and automated jobs that stay connected to the network. That led the company to ask a different question: what would version control look like if agents were assumed to be online all the time from the start?
Three ways Scion differs from Git
Anthropic said Scion changes the model in three main ways.
- First, the client is assumed to be always online, so it does not keep the entire repository locally. It fetches only the data it needs from the server.
- Second, heavy operations move to the server. Line blame, commit creation, and merges are handled there. In Git, changes are committed locally and then pushed. In Scion, there is no push step.
- Third, many tasks do not require code to be checked out at all. Queries used by clients are exposed through APIs, so if CI needs to compare two versions, it can ask the server directly.
Anthropic added that there is no single complete repository on the server side either. Data is stored in a distributed way, and each component can scale independently. Code access goes through a three-layer cache.
Performance numbers Anthropic shared
According to the company, every cache layer in Scion records a hit rate above 99%. It also said the CPU and memory required for Scion’s infrastructure are 200 times smaller than what its earlier fleet of Git mirror servers needed. CI and research workloads now check out code two orders of magnitude faster, falling from minutes to seconds. CI alone saves 25% of CPU capacity, Anthropic said.
Anthropic says most companies may not want to use it directly
The company also acknowledged the limits of the design. It said these choices only make sense for a very large monorepo, and that Anthropic can spend far more tokens internally than most companies would be willing to spend. Because Scion differs so much from Git, adopting it would mean replacing much of a company’s development pipeline and pairing it with its own code review, search, and CI systems. Anthropic said that was one reason it did not want to force an external version built to satisfy everyone.
Even so, the company expects these problems to become common over the next six to 12 months. That is why it chose to publish the architecture, with two practical suggestions. One is that Git servers do not need to operate the same way as Git clients. The other is that hosting platforms such as GitHub could provide a more complete set of server-side APIs like the ones Scion uses.
Pressure is building across the Git ecosystem
Anthropic is not alone in facing this shift. BlockTempo previously reported that GitHub recorded 7.38 billion commits in September, more than five times the level from a year earlier, and has decided to rewrite Git’s storage layer.
Reaction to Scion has been mixed. Anthropic said the post had reached 885 points and about 1,000 comments on Hacker News as of Oct. 16. A common criticism was that the system amounts to rebuilding older centralized version control tools such as Perforce and Subversion.
Anthropic itself cited prior examples in the post, including Google’s Piper and Meta’s Sapling. In an Oct. 15 update, the company added that Sapling’s server-side system, Mononoke, pioneered many of the techniques Scion now uses.
The most-replied-to comment on Hacker News focused on code review. The commenter said that whenever Claude-written code goes unreviewed, “garbage” tends to show up later, and asked how Anthropic avoids that problem. As of publication, neither GitHub nor GitLab had publicly responded, according to the article.