Anthropic Replaces Git with In-House VCS 'Scion' as AI Merges Explode

As Claude generated an unprecedented volume of code, Git hit scaling walls first—prompting Anthropic to build and migrate its massive monorepo to ‘Scion’, a server-centric in-house version control system.

Anthropic published an engineering post in October detailing Scion, its proprietary version control system, and Antique, a dedicated code review platform. Faced with rapid agentic code generation, a team of seven engineers built Scion and fully migrated the company’s monorepo from Git in five months.

Why Git broke down

The catalyst was an explosion in code written by AI agents (programs that carry out multi-step tasks for you on their own). In May 2026, Anthropic averaged roughly 2,500 merged changes per day. By September, merges surged past 30,000 per day—a twelvefold jump in four months.

The repository outgrew human scale. The main branch history approached 100 GB, with compressed working copies topping 25 GB. Continuous integration (CI) alone ran around 25 million checkouts every day, before counting developer environments, agent sandboxes, and research jobs. Pulling that naively via Git would have generated more than 600 PB of network traffic daily.

Anthropic initially scaled with several hundred internal Git mirrors and repository seed snapshots. But as merge frequency escalated, mirrors fell behind main or ran out of memory, repeatedly disrupting developer and agent workflows.

How Scion departs from Git

Git is fundamentally decentralized, designed around offline-capable clients holding a full repository clone. In Anthropic’s workload, however, almost every client reading and writing code is an always-online AI agent or automated CI job.

Scion rearchitected source control around always-connected clients, drawing on concepts from Google’s Piper and Meta’s Sapling. Instead of cloning the full repository, clients use APIs to fetch only the specific directory trees and file blobs needed for a task.

Commits record only the changesets relative to parents in a Spanner database. Operations like file blame and history are computed on demand as derived data rather than traversing local packfiles. Node daemons and regional caches stream objects directly from memory, maintaining cache hit rates above 99%.

The resulting infrastructure requires 1/200th of the CPU and memory footprint of the previous Git mirrors. Code checkout times dropped from minutes to seconds, and CI systems recovered 25% of their CPU capacity.

Rapid migration and clear limits

Seven engineers built and deployed the system in five months, aided significantly by Claude itself. For example, when an early snapshot of Claude rewrote the handwritten Git-to-Scion ingestion prototype, it cut memory consumption by 20x and quadrupled processing speed while preserving data invariants.

Anthropic emphasized that Scion is not a general-purpose Git replacement. It is not open source, offers no offline support, and is tightly coupled to Anthropic’s specific monorepo, internal code search, and the Antique review tool.

What to watch next

Anthropic expects similar scaling bottlenecks to confront other engineering teams as autonomous coding agents spread over the next 6 to 12 months. Rather than abandoning Git entirely, the broader ecosystem is likely to explore hybrid patterns—such as richer server-side RPCs, scalable cache daemons, and improved sparse checkout protocols.

Sources

• Redesigning source control at Anthropic (Anthropic Research)