TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Anthropic says it built Scion, an in-house version control system, and migrated its monorepo from Git in five months. The company reports that its daily merged code changes rose from about 2,500 in May 2026 to more than 30,000 in September, putting pressure on infrastructure designed around conventional development workloads.
Anthropic says it built an in-house version control system called Scion and moved its monorepo from Git to the new system in five months, responding to a sharp rise in code changes and automated repository workloads. The company reported averaging about 2,500 merged changes a day in May 2026 and more than 30,000 a day in September.
In its engineering report, Anthropic said the pressure point had shifted from writing code to the infrastructure needed to build, review and merge it. The company attributed the change to Claude’s ability to work through research and coding tasks at greater speed and scale. That explanation is Anthropic’s account of its own workload, not an independently measured assessment of AI coding across the industry.
The report describes a repository whose main-branch history is nearly 100 GB, with a compressed working copy exceeding 25 GB. Anthropic said it runs roughly 25 million code checkouts per day for CI, before counting developer environments, agent sandboxes and research jobs. A naive Git-based approach to fetching those copies would amount to more than 600 petabytes of traffic daily, the company estimated.
Before building Scion, Anthropic used Git scaling techniques including shallow clones, sparse checkouts, internal mirrors and repository snapshots stored in blob storage. It said it operated several hundred internal Git mirrors at peak. As merges increased, the company found the setup harder to maintain: snapshots could fall behind the main branch, while operations such as diff and blame became more complicated on partial clones.
Why Scion Fits Anthropic’s Workload
Scion reflects a trade-off between local independence and centralized services. Git is designed so a clone contains repository data needed to inspect history and work offline. Anthropic said that model was increasingly costly for its own use, where most clients are online agents and jobs that repeatedly need files or repository information.
Under Scion’s design, the server handles operations such as answering history and blame queries, creating commits, and performing merges and rebases. Clients keep a bounded local cache, arrange working files on disk and report filesystem changes. Anthropic says this reduces the amount of repository data each client needs to hold. It also makes source-control queries available through an API, which the company says helps automated jobs and internal tools avoid maintaining full local copies.
The change matters beyond repository storage because source control is part of the automation pipeline. If code generation produces more changes, review, integration and CI capacity can become constraints. Anthropic’s report presents Scion as one element of a broader infrastructure effort that also included a code-review interface called Antique, code search and access-control systems. The company’s unusually high workload limits how directly other organizations can apply its results, but the design choices may be relevant as automated coding increases demand elsewhere.
version control system for software development
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Git Mirrors to Scion
Git’s distributed design lets each clone operate as a complete repository, including access to history and offline work. Anthropic said that approach remains appropriate for most repositories. Its argument is narrower: a very large monorepo, served to many online agents and jobs, faces a different balance of costs than a typical developer repository.
The company said standard Git features and internal workarounds initially helped it handle growth, but the number of merges made mirror and snapshot operations harder to manage. It then chose to build around clients that could reach a server, rather than continuing to provide each client with a largely self-contained copy.
Anthropic’s report cites Google’s Piper and Meta’s Sapling as prior work on source control for large monorepos. It says those systems provided lessons for Scion, which was developed alongside other source-control tools during the five-month migration. The report focuses on Scion’s architecture rather than presenting a full, independent comparison of the systems.
“The bottleneck, now, is all of the developer infrastructure built for human scale: source control, builds, and CI.”
— Anthropic, in its engineering report
high capacity external hard drive for code storage
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What the Report Does Not Measure
The report does not provide an independent performance evaluation comparing Scion with Git, nor does the supplied material quantify the new system’s operating costs, reliability, migration risks or impact on developer productivity. The reported merge counts and checkout volumes are Anthropic’s figures; the material does not include an external audit or methodology for those measurements.
It is also unclear how much of Scion’s design or tooling will be released outside Anthropic, and whether other companies could adopt it without similar infrastructure and workloads. Anthropic says its experience is unusual and does not claim that most organizations need to replace Git now. The company’s expectation that AI-related repository pressures will spread is a projection, not a confirmed industry-wide trend.
As an affiliate, we earn on qualifying purchases.
Further Scion Design Details
Anthropic’s report sets out the motivation and principles behind Scion, including online clients, a division of work between server and client, API access to repository queries, and horizontal scalability. The supplied material ends while discussing scalability and does not establish a public release date, adoption plans outside the company or a schedule for further technical disclosures.
The next useful evidence will be operational detail: how Scion performs under Anthropic’s reported workload, how it handles outages and growth, and how its tools support reviews and CI in practice. Until the company publishes those details or outside evaluations become available, Scion’s results should be understood as Anthropic’s account of an internal system built for its own needs.
software development automation tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Scion?
Scion is Anthropic’s in-house version control system, built to manage the company’s monorepo. Anthropic says its server handles repository operations while clients keep a smaller local cache and working copy.
Why did Anthropic move away from Git for its monorepo?
Anthropic said the monorepo’s size and volume of automated checkouts made its Git mirrors and snapshots harder to operate as the company merged more code. It designed Scion around the fact that most of its clients are online agents and jobs.
How many changes does Anthropic say it merges each day?
Anthropic reported 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.
Does Anthropic say other companies should replace Git?
No. The report says Git remains the right design for most repositories and describes Anthropic’s workload as unusual. The company suggests similar pressures may reach more organizations as AI coding tools are used more widely, but that is a prediction.
Is Scion available to the public?
The report describes an internal migration but does not say whether Scion will be released publicly or adopted outside Anthropic.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
