Redesigning Source Control At Anthropic

TL;DR

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Anthropic says it built Scion, an in-house version control system, and migrated its monorepo from Git in five months. The company reported more than 30,000 merged code changes per day in September 2026, up from an average of about 2,500 in May, and says its agent-heavy workload strained infrastructure designed for smaller, human-scale development.

Anthropic has replaced Git for its monorepo with Scion, an in-house version control system the company says it built and deployed in five months. The redesign responds to a rapid rise in code output: Anthropic reports that its daily merged changes increased from about 2,500 in May 2026 to more than 30,000 in September, as AI agents and research jobs placed new demands on developer infrastructure.

Anthropic described the change in an engineering report focused on Scion, the system it built to manage source code. Alongside Scion, the company says it developed a code-review interface called Antique, code search, an internal source-control access-control system, and new ways to manage and land large volumes of reviewed code. The report says those tools were built and shipped during the same five-month period, though it focuses mainly on Scion.

The company’s monorepo had become difficult to serve using a conventional Git setup. Anthropic says the main branch’s history is nearly 100 GB, while a compressed working copy exceeds 25 GB. It reports about 25 million code checkouts per day for continuous-integration jobs alone. A naive full-fetch approach, it estimates, would generate more than 600 petabytes of Git traffic daily; that is a hypothetical estimate, not a measurement of actual traffic.

Anthropic says it had relied on Git features such as partial clones and sparse checkouts, as well as internal mirrors and stored repository snapshots. At peak, it operated several hundred Git mirrors. The report says those workarounds became harder to manage as the merge rate rose, with snapshots falling further behind the latest code and history-dependent operations such as blame becoming harder to support on partial clones.

Why Anthropic Moved Beyond Git

The change reflects a mismatch between Git’s distributed design and the way Anthropic says much of its code is now produced and consumed. Git gives each clone a substantial local repository, allowing developers to inspect history and work offline. Anthropic says most of its relevant clients are instead online agents, CI jobs, and research workloads. For those users, repeatedly distributing repository data can be expensive, while offline operation is less central.

Scion shifts more work to a source-control server. According to the report, the server handles tasks such as answering history and blame queries, creating commits, and performing merges and rebases. Clients keep a bounded local cache, manage working files, and report file changes. Anthropic argues that this arrangement requires less local data and can make repository information available through APIs, so some automated jobs can query source control without maintaining a full local checkout.

The practical stakes extend beyond the version-control tool itself: Anthropic says coding is no longer its main development bottleneck, while systems for source control, builds, and CI are under pressure. The company presents its experience as an early warning for organizations that use AI coding tools at high volume. That is a prediction based on its own workload, not evidence that other companies will face the same scale or need the same architecture soon.

From Git Mirrors to Scion

Git’s decentralized model is well suited to projects where developers benefit from complete local copies and offline work. Anthropic says it initially used Git’s large-repository options and conventional workarounds rather than immediately replacing it. Those included shallow and sparse checkouts, internal mirrors, and repository seeds stored in blob storage. The company says the setup bought time but grew harder to operate as more changes were merged.

Anthropic frames Scion around three principles in the report: clients are expected to be online; the server and client have different responsibilities; and the system should scale horizontally. It also cites earlier large-scale source-control systems, including Google’s Piper and Meta’s Sapling, as prior work whose lessons informed its approach. The account is Anthropic’s own engineering report; it does not provide an independent evaluation or a head-to-head performance comparison with Git.

“The bottleneck, now, is all of the developer infrastructure built for human scale: source control, builds, and CI.”

— Anthropic, in its engineering report

What the Report Leaves Open

The report is a description by the system’s creator, and it does not include independent performance data, detailed cost comparisons, or a direct benchmark against the previous Git setup. It also does not specify Scion’s availability targets, how it handles prolonged network outages, or how the system’s operational demands compare with maintaining Git mirrors.

Anthropic’s numbers describe its own development environment. The company says it merged over 30,000 changes per day in September 2026, but the report does not define every detail of what counts as a change or provide a comparison baseline for the daily checkout figure. It remains unclear whether Scion will be made available outside Anthropic or whether other organizations with heavy AI-assisted coding workloads would benefit from a similar design.

How Scion Performs at Scale

Anthropic’s report sets out the architecture and reasons for the migration, but it does not announce a public release or a rollout to other companies. The next evidence to watch for is how Scion performs as the company’s code volume and automated workloads grow, including whether its server-based queries and horizontally scalable services continue to meet demand.

The company also identifies builds and CI as parts of the broader infrastructure bottleneck, but the supplied report does not describe a specific next project or timeline for those systems. Any further plans, external availability, and results from operating Scion over a longer period remain unannounced in the material provided.

Key Questions

What is Scion?

Scion is Anthropic’s in-house version control system, built to serve the company’s monorepo. Its server handles operations such as history queries, commits, merges, and rebases, while clients keep a bounded local cache and manage working files.

Why did Anthropic move away from Git for its monorepo?

Anthropic says its repository’s size and high volume of automated checkouts made its Git mirror and snapshot setup increasingly difficult to operate. It designed Scion around online agents and jobs that can request source-control data from a service instead of carrying a full repository locally.

How much did Anthropic’s code merges increase?

Anthropic reports an average of about 2,500 merged changes per day in May 2026 and more than 30,000 per day in September 2026. These are company-reported figures; the report does not give further details about how a change is counted.

Did Anthropic make Scion available to other companies?

The report describes Scion as an internal system and does not announce a public release or availability for other organizations.

Does the report prove Scion is faster than Git?

No direct comparative benchmark is included in the supplied report. Anthropic explains why it built Scion and describes its architecture, but independent performance results and detailed cost comparisons are not provided.

Source: hn