Anthropic Replaces Git With Scion as Code Changes Surge

Anthropic has replaced Git with an internally developed version control system called Scion after a surge in AI-generated software overwhelmed its existing development infrastructure. The company completed the migration of its monorepo in five months, following a sharp increase in daily code merges from approximately 2,500 in May 2026 to more than 30,000 in September 2026. In a technical article published on October 13, 2026, Anthropic explained that its growing reliance on Claude for software development had shifted the main constraint from writing code to managing source control, builds, and continuous integration (CI).

The company developed Scion alongside other internal tools to accommodate its expanding workload. Anthropic also built a code review interface called Antique, code search capabilities, an identity and access management system for source control, and mechanisms for reviewing and integrating large volumes of code. These systems were developed during the same five-month period as the migration.

The company said it does not expect demand for code integration to slow soon, as removing infrastructure constraints enables further increases in software production.

Anthropic's monorepo had grown beyond the scale that its existing Git infrastructure could efficiently support. According to the company, the history of its main branch alone approached 100 GB, while a compressed working copy exceeded 25 GB. The company also performed approximately 25 million code checkouts daily for CI, excluding developer environments, agent sandboxes, and research jobs. Anthropic estimated that downloading repository data naively for these operations would generate more than 600 PB of Git traffic per day.

To manage the load, Anthropic previously used several approaches available in Git, including shallow clones, sparse checkouts, and promisor remotes. It also operated several hundred internal Git mirrors at peak capacity and maintained repository snapshots in blob storage. These measures initially helped accommodate growth, but became increasingly difficult to maintain as merge volumes rose. Repository snapshots fell further behind the latest main branch between refreshes, while operations such as comparing changes and identifying the last author of a line became harder to maintain on partial clones.

Anthropic concluded that Git's decentralized design was poorly matched to its increasingly centralized workload. Although Git allows every clone to retain a complete repository and operate offline, most of Anthropic's development clients were agents and automated jobs that remained connected to the network. Scion was designed around that difference.

Rather than requiring every client to maintain a complete repository, the system relies on an accessible server that supplies information and performs operations when requested. Under this architecture, the server handles file history queries, blame information, commit creation, merges, and rebases. Clients maintain a limited local object cache, manage working files, monitor filesystem changes, and provide developers with an interface for interacting with the server.

Anthropic said this division reduces the amount of repository data that individual clients need to store. It also makes source control functions available through an API, allowing automated jobs and internal tools to retrieve repository information without maintaining a full local copy.

The company cited automated bisect and revert pipelines as examples of tools that benefit from inexpensive programmatic access to source control. These capabilities allow internal systems to investigate code changes and reverse them when necessary without requiring every job to operate on a complete repository. Scion's design also emphasizes horizontal scalability, with repository data distributed rather than concentrated in individual mirrors.

Anthropic said its experience operating hundreds of Git mirrors had exposed capacity limits when repositories exceeded available disk space or memory. Anthropic noted that large-scale source control systems already have precedents, including Google's Piper and Meta's Sapling. The company developed Scion with lessons from these systems in mind, while adapting its architecture to an environment where AI agents and automated jobs account for a substantial share of software development activity.