Anthropic has unveiled an internal version control system called Scion, after its Git-based architecture and repository mirrors began struggling to keep pace with the rapidly accelerating volume of software changes produced by AI tools. The company says daily code merges increased from 2,500 in May 2026 to more than 30,000 in September of the same year.
Anthropic links this increase to the intensive internal use of Claude for software development. As each bottleneck is removed, the volume of code produced rises again, making development infrastructure tools designed around the pace of human work—including source control, builds, and continuous integration—the slowest element in the process.
A System Designed for a Massive Repository
According to the company, the infrastructure footprint required was reduced by up to 200 times in terms of processor and memory consumption compared with the previous architecture, which relied on mirroring Git repositories. CI operations and research tasks can also now fetch code at speeds two orders of magnitude higher.
Anthropic migrated fully to Scion within five months after it was developed by a team of seven people. During the same period, the team created a new code review tool called Antique and approximately six other systems, according to the company.
How Does It Differ from Git?
Scion operates on assumptions different from those of Git. It provides continuous access to the server instead of assuming that a client needs a complete copy of the repository to work offline. It also offers an API that CI tools can query for the code history and repository information without maintaining a local repository.
Anthropic says the system scales horizontally, with repository data distributed across components that can be scaled independently, reducing the need to manage hundreds of Git mirrors. According to the company, it also underwent simulation and fuzzing tests, along with advanced rate-limiting controls.
Why Does This Matter?
Scion is not a ready-made general-purpose replacement for Git, but it reveals the kinds of pressures that coding agents capable of producing software changes at a pace exceeding the design of traditional development tools may impose. The practical impact is not limited to the source control system; an increase in the volume of changes can shift the bottleneck to builds, CI, or review.
Anthropic expects other teams to face the same problem as AI-based software development expands, but it does not plan to release Scion publicly. It explains that the system is highly specialized and that operating it could require substantial spending on tokens, in addition to migration costs resulting from its fundamental departure from Git. Instead, the company will seek to share some of its ideas with GitHub, GitLab, and the open-source community.
Limitations and Open Questions
Developers outside Anthropic do not have access to Scion’s code or to a public product they can test, and the company has not made any commitments to release a public version. Based on the available information, the announcement’s value lies in presenting an internal infrastructure experiment and performance indicators, not in providing a new tool to the market. The open question remains whether ideas such as persistent access, APIs, and horizontal scaling will transfer to Git or other public systems without the migration costs and complexity Anthropic warned about.