Why Anthropic Replaced Git for AI Development

AI leader Anthropic has replaced Git with a custom source control system called Scion to better manage its massive, single code repository.

AI research company Anthropic has taken the significant step of replacing Git, the industry-standard source control system, with a custom-built tool named Scion. According to a report from InfoQ, the company migrated its large, unified codebase, often called a monorepo, to the new system to better handle the unique scale and speed required for modern AI development. Unlike Git, which operates on a distributed model where every developer has a full copy of the repository's history, Scion centralizes operations on a server. This design choice represents a fundamental trade-off: developers lose the ability to work completely offline, but in return, they gain significant performance improvements by not having to manage a massive repository on their local machines. The move highlights a growing challenge in the tech industry, where foundational developer tools are being pushed to their limits by the demands of large-scale software and AI projects.

The core innovation of Scion lies in its server-centric architecture. By shifting the heavy lifting of repository operations from individual developer laptops to powerful central servers, the system drastically reduces the amount of data that needs to be downloaded and processed locally. Scion uses a technique called content-addressed storage, which is similar to how Git stores objects, to efficiently manage data and avoid duplication across the entire system. To further boost performance, especially for automated processes like continuous integration (CI), it employs tiered caches. This means frequently accessed code and files are stored in faster, more readily available memory, allowing automated build and test systems to run much more quickly. For developers, this translates to faster commands and a more responsive workflow, even when dealing with a repository that contains vast amounts of code, data, and history.

Anthropic's decision is not without precedent among technology giants. Companies like Google and Meta have also developed their own internal source control systems, Piper and Sapling respectively, to manage their colossal monorepos that outgrew Git's capabilities years ago. Git, for all its strengths in the open-source world and for small-to-medium projects, can struggle with repositories containing millions of files and a deep, complex history. The rise of AI development adds another layer of complexity, often involving large model files, datasets, and intricate dependencies that further strain traditional version control. Anthropic's move is part of a broader trend where hyper-scale companies find it necessary to build bespoke infrastructure to optimize developer velocity and maintain a competitive advantage. It's a classic build-versus-buy decision applied to one of the most fundamental tools in the software development lifecycle.

For most engineering teams, replacing Git is not a practical or necessary step. However, Anthropic's experience offers a valuable lesson for any organization grappling with performance issues in a large monorepo or a complex AI development environment. It underscores the importance of critically evaluating whether standard, off-the-shelf tools are still the right fit as workloads evolve and scale. The key takeaway is not to abandon established tools, but to be proactive in identifying bottlenecks in the development process and exploring alternative architectures. The industry should now watch to see if other AI labs follow Anthropic's lead, potentially creating a market for new commercial or open-source source control systems designed specifically for the AI era. This could signal the beginning of a fragmentation in a tool category that Git has almost completely dominated for more than a decade.

Anthropic's move away from Git signals that standard developer tools may not scale for the unique demands of AI development. This case study challenges infrastructure teams to re-evaluate their source control strategy, especially for large monorepos and intensive CI/CD pipelines common in machine learning.

By building a custom source control system, Anthropic is betting on long-term developer productivity gains that outweigh the cost of creating bespoke infrastructure. This move could provide a competitive edge by accelerating AI research and development, pressuring rivals to optimize their own tooling.