October 13th 2026, 4:40 PM
Anthropic Replaces Git with Custom Source Control System Scion to Handle 30k Daily Code Merges
Key Summary
Anthropic has built a custom source control system called Scion to handle its rapid code merges, driven by its agentic coding tool Claude. The new system replaces Git and boasts significant performance improvements, including faster merge times and reduced costs. This move highlights the need for tailored version control and CI tooling in the AI industry.
Please see our real time news feed on our Home Page
Background
Anthropic has replaced Git with a homegrown source control system called Scion.The Need for Scion
In May 2026, Anthropic averaged roughly 2,500 merged code changes per day. By September 2026, that figure had climbed past 30,000 merges per day.The Rise of Claude
The company attributes the surge to the growing capabilities of its Claude models. Claude increasingly handles coding tasks on its own, which means code is being produced at a rate that looks less like a team of engineers and more like a factory floor.Designing Scion
Seven engineers, five months Rather than patch the old setup, Anthropic built a new one. A team of just seven engineers designed Scion and completed the full migration of the company's monorepo in five months.Key Features of Scion
Scion borrows design ideas from two systems built by companies that faced similar scale problems: Google's Piper and Meta's Sapling.Performance Improvements
Scion uses always-online clients, meaning the developer's tools stay connected to central servers rather than holding a full independent copy of everything. Scion also moves commits and merges to the server side. CI job completion times dropped from minutes to seconds under Scion.Benefits and Challenges
Scion reaches cache hit rates above 99%. Its CPU and memory footprint is 200 times smaller than the Git mirror infrastructure it replaced.What This Means for Software Teams
There are risks too. Moving away from Git means leaving behind a massive ecosystem of compatible tools, integrations, and developer familiarity. Custom infrastructure also requires ongoing maintenance, and its long-term costs are harder to predict than its launch-day benchmarks.#AI#US#Crypto#SoftwareDevelopment