Anthropic says it moved its main software repository from Git to an in-house system called Scion in five months, after daily code changes merged into the repository rose from about 2,500 in May to more than 30,000 in September 2026.
The change addresses a less visible consequence of AI coding agents: at very high volumes, the infrastructure that stores code, checks it and combines changes can become a bottleneck. Anthropic's engineering team described the migration on October 13. Its results are company-reported; the post is not an independent benchmark or a public release of Scion.
A monorepo is one code repository shared by many software projects or services. Git is a distributed version-control system: its standard clone gives each user a copy of the repository and its history, a design that lets developers work independently and offline. That design served Anthropic well until the company's repository and automated workloads reached unusual scale.
Scion shifts repository work to online services
Anthropic says the history of its main code branch is almost 100 gigabytes, while a compressed working copy is more than 25 gigabytes. Its continuous-integration (CI) systems, which automatically build and test proposed code changes, perform about 25 million code checkouts a day. That count excludes developer computers, agent workspaces and research jobs. Anthropic estimates that naively transferring a full Git repository for those checkouts would exceed 600 petabytes a day.
The company had run several hundred internal Git mirrors to handle the load, but says they became harder to operate as the repository and merge rate grew. Snapshots used to seed new checkouts fell further behind the main branch between refreshes, and partial copies made history-dependent tasks such as blame harder to maintain.
Scion assumes that most clients, including AI agents and CI jobs, are online and can ask a server for what they need. The server answers history questions and performs commits, merges and rebases; clients keep a bounded cache and request selected files as needed. Sparse checkouts contain only requested directories, while virtual checkouts fetch files on demand. Some jobs can query repository history or request a diff through an API without creating a full checkout.
Anthropic reports that Scion uses 200 times less CPU and memory than the former Git-mirroring setup. It says CI and research jobs check out code about 100 times faster, and CI gained 25% more CPU capacity. The company also reports cache hit rates above 99% across Scion's tiers. These figures describe Anthropic's own workload and comparison; the post does not publish an independent test or a breakdown across other organizations.
A specialized system, not a general Git replacement
Anthropic says a seven-person team built Scion, its code-review tool Antique and related systems, then moved the monorepo over in five months. Engineers still reviewed every change request, the company says, but focused on the intent of each change rather than reading every line. Because Scion does not use a global lock to serialize all merges, the team relies on testing before deployment and after changes land to catch conflicts and failures.
The tradeoffs matter. Anthropic says Scion is not suitable as a general replacement for Git: it is aimed at extremely large monorepos or environments with many always-online agents, and it requires its own code-review, search and CI systems. Anthropic has not announced Scion as a public product.
Sources
- Anthropic Engineering: Redesigning source control at Anthropic (published October 13, 2026).
- Pro Git: About Version Control (how distributed systems such as Git provide clients with repository history).
- Illustrative image: “Datacenter Rack” by Edmondo, licensed under CC BY-SA 3.0; image unchanged.
