Anthropic Replaces Git With Custom Source Control System 'Scion'
Anthropic has migrated its codebase from Git to a new, in-house centralized source control system called Scion. The move addresses extreme scale requirements arising from a dramatic increase in code changes after recent model releases. Scion sacrifices offline support in favor of high throughput and CI integration, primarily serving automated and agentic workflows.
What changed?
Anthropic has replaced Git as its version control backbone with a custom system named Scion. Developed by a small engineering team over five months, Scion is modeled as a centralized alternative, similar to Subversion. The shift was driven by the company’s need to handle 30,000 code changes merged per day, manage a main branch history of nearly 100 GB, and support more than 25 million code checkouts daily for CI and automated workflows.
Why does it matter to an everyday developer?
Anthropic’s migration showcases the limits of even mature systems like Git in the context of rapid, large-scale, and AI-driven development. The company’s core workflows rely on automated agents such as Claude, which require fast, online-first and high-frequency repository access—something not addressed efficiently by standard distributed version control. The absence of offline support in Scion is a deliberate tradeoff, acknowledging that the heaviest consumers of code are automated systems, not humans. This example highlights a growing divergence between conventional human-centered tooling and workflows built for and around LLMs and automated agents.
What can the developer do now?
Most developers will not interact directly with Scion, as it is an internal solution tuned for Anthropic’s scale and agentic requirements. However, this move signals that developers working in environments with extreme CI or automation needs should be aware of the limitations of mainstream tools such as Git. Teams supporting AI/LLM-driven workflows may want to reassess their infrastructure and consider whether a centralized model or custom solutions are warranted, especially when agent usage vastly outweighs human interaction. Observing Anthropic’s experience may inform future architecture decisions for tool builders facing similar scale bottlenecks.
