Anthropic Replaces Git With Custom Scion System
Anthropic replaced Git with a custom internal system named Scion after automated AI coding pushed their daily code merges from 2,500 to over 30,000 in just four months.

Anthropic has abandoned traditional Git version control in favor of a bespoke centralized system called Scion, paired with a custom code review tool named Antique. The shift was driven by an unprecedented surge in automated software development within the company. Between May 2026 and September 2026, Anthropic saw its daily merged code changes explode from roughly 2,500 to more than 30,000. This spike was largely propelled by the internal deployment of advanced AI models, including its Fable and Mythos classes.
The technical scale of Anthropic's codebase made standard distributed version control impractical. The history of the main branch alone reached nearly 100 GB, with a single compressed working copy exceeding 25 GB. Across continuous integration pipelines, research tasks, developer setups, and autonomous agent sandboxes, the system executes around 25 million code checkouts every day. Transporting that workload over conventional Git protocols would have consumed more than 600 PB of daily traffic.
A team of seven engineers built Scion and completed the migration from Git in five months. Because the primary user of the repository is the Claude AI model itself—which operates online—the developers intentionally omitted offline support. Anthropic also relies heavily on adversarial code reviews executed by Claude, though human engineers must actively guide the model to prevent it from over-engineering solutions for hypothetical issues.
For engineering leaders and infrastructure practitioners, Anthropic's transition highlights a fundamental shift in software engineering tools. As autonomous AI agents become the dominant authors of codebases, traditional version control systems designed for human velocity and offline flexibility are increasingly giving way to specialized, high-throughput architectures built for machine-scale workflows.
This is our own summary of reporting by Simon Willison



