Redesigning Source Control At Anthropic
AIThis post was created with the assistance of artificial intelligence (AI).

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
As an affiliate, we earn on qualifying purchases.

Anthropic says it built Scion, an in-house source-control system, and moved its monorepo off Git in five months. The company reported that daily merged changes rose from about 2,500 in May 2026 to more than 30,000 by September, putting pressure on its existing developer infrastructure.

Anthropic says it built Scion, an in-house source-control system, and migrated its monorepo from Git in five months after a rapid increase in code production strained infrastructure designed for human-scale workloads. The company reported that its average daily merged changes rose from about 2,500 in May 2026 to more than 30,000 by September.

Scion changes how repository data is handled. Rather than giving every client a complete repository copy, as Git’s decentralized design allows, Scion relies on an always-reachable server for operations including history queries, blame, commit creation, merges and rebases. Its clients keep a bounded local object cache, maintain working files and report filesystem changes, while fetching data as needed.

Anthropic says the system was developed alongside other internal tools, including a code-review interface called Antique, code search and a source-control permissions system. The report focuses on Scion, describing its design principles as a response to a workload in which many repository users are online agents and jobs rather than people working offline.

The scale figures in the report help explain the change. Anthropic says the history of its main branch is nearly 100 GB, a compressed working copy exceeds 25 GB, and its CI system performs about 25 million checkouts a day, not counting developer environments, agent sandboxes or research jobs. The company estimates that naive Git fetching at that checkout volume would generate more than 600 petabytes of traffic per day.

At a glance
reportWhen: Anthropic’s report describes a five-mon…
The developmentAnthropic says it created Scion and migrated its monorepo from Git in five months to handle a sharp increase in code changes and repository workloads.

Why Anthropic Changed Its Source Control

The project reflects a growing infrastructure challenge for organizations that use coding agents at high volume: faster code production can shift limits from writing software to review, builds, CI and source control. Anthropic says that removing one bottleneck quickly led to demand for more code merging, making repository operations a constraint on its internal work.

Scion’s server-centered design trades some of Git’s offline independence for less data on each client and centralized access to repository operations. That can matter for automated systems that need to inspect history or perform repository tasks without repeatedly maintaining full local copies. Anthropic says its internal automated bisect and revert pipelines benefit from programmatic source-control access, though the report does not provide quantified performance comparisons.

The design is not presented as a general replacement for Git. Anthropic describes its own workload as unusual and says most repositories remain well suited to Git’s decentralized model. The report’s wider relevance is that agent-heavy development could lead some organizations to revisit infrastructure assumptions as their workloads grow.

Amazon

source control management software

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

From Git Mirrors to Scion

Git gives each clone a full, self-contained repository, allowing users to inspect history, compute diffs and check out revisions while offline. Anthropic says that model became increasingly costly for its monorepo as the repository and checkout volume grew. It had tried Git’s partial-repository features and operational workarounds, including internal mirrors and stored repository snapshots known as seeds.

At peak, according to the report, Anthropic operated several hundred Git mirrors. The company says rising merge rates made those systems harder to manage: snapshots could fall further behind the main branch between refreshes, and partial clones required careful handling for history-dependent features such as diff and blame.

Anthropic says it took lessons from earlier large-scale source-control systems, including Google’s Piper and Meta’s Sapling. It then designed Scion around three stated principles: clients are online, server and client responsibilities are asymmetric, and the system must scale horizontally. Its report presents the system as an in-house response to its particular needs, not as a claim that other companies should immediately migrate.

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

— Anthropic, in its engineering report

Scion’s Results Beyond Anthropic

The report does not provide a detailed benchmark comparing Scion with Git, nor does it quantify the system’s impact on checkout speed, reliability, operating costs or developer productivity. It also does not specify the exact dates of the five-month migration or describe whether every repository and workflow at Anthropic now uses Scion.

Anthropic’s figures and conclusions are company-reported. The material available here does not include independent verification, details of Scion’s availability to outside users, or enough information to assess how its server dependence performs during outages or connectivity loss. Whether the same approach would benefit other organizations depends on their repository size, workload and need for offline work.

How Anthropic Will Extend Scion

Anthropic’s report outlines the system’s architecture and the pressures that led to it, but does not announce a public release, an external adoption plan or a specific next milestone. The immediate next steps described are operational: Scion and related tools must support the company’s continuing source-control workload as its code production grows.

Readers can look for further technical detail or operational results from Anthropic, including whether it publishes performance measurements or discusses lessons from running Scion at scale. Until then, the reported migration and workload figures remain the clearest account of what the company has built and why.

Key Questions

What is Scion?

Scion is Anthropic’s in-house version-control system. The company says its server handles repository operations such as history queries, commits, merges and rebases, while clients keep a limited local cache and working copy.

Why did Anthropic move away from Git for its monorepo?

Anthropic says the size of its monorepo and the volume of automated checkouts made its Git mirror and snapshot setup harder to operate. It designed Scion for clients that are generally online and can fetch repository data when needed.

How quickly did Anthropic’s merged changes increase?

Anthropic reported an average of about 2,500 merged changes per day in May 2026 and more than 30,000 per day by September 2026. These are company-reported figures for those periods.

Is Scion available to other companies?

The report does not announce a public release or say that Scion is available to outside users. Anthropic describes it as a system built for its own monorepo and workload.

Source: hn

HALLOWEEN

Halloween Picks

As an affiliate, we earn on qualifying purchases.

You May Also Like

Layer‑Zero Messaging: The Next Frontier in Interoperability

Introducing Layer-Zero Messaging, the innovative approach that could transform interoperability by enabling secure, trustless data exchange between blockchains—discover how it works.

Show HN: Whiteboard (YC W26) – An Open-source IDE For Thoughtful Software Design

Whiteboard is an open-source desktop IDE aimed at enhancing software design, created by YC W26 founders. Its development signals interest in thoughtful coding tools.

Exploring Claude/GPT Knowledge Cutoffs And Pre-Training Timelines

An analysis of the knowledge cutoffs and pre-training timelines for Claude and GPT models, highlighting confirmed facts and ongoing uncertainties.

The Game-Changing AI Settings That Tripled Our ARC-AGI-3 Performance

OpenAI reports that enabling two unspecified settings on its model tripled performance on the ARC-AGI-3 benchmark, highlighting evaluation sensitivity.