Anthropic replaces Git with in-house version control
With Claude driving daily merges to 30,000, Anthropic replaced Git with Scion, an online version control system built for its massive monorepo.
Driven by Claude agents pushing 30,000 daily merges, the company built Scion in five months to support its 100 GB monorepo.
Anthropic has replaced Git across its main monorepo with an in-house version control system named Scion. The company’s engineering team announced the rollout on Tuesday, October 13, 2026. The decision came after Claude models began writing and submitting changes so rapidly that traditional development infrastructure buckled under the load.
Between May and September 2026, average daily merges into Anthropic’s repository surged from 2,500 to over 30,000. That twelvefold increase was driven almost entirely by AI agents writing, committing, and submitting code around the clock.
Why Git broke under agent workload
Git was designed in 2005 for human developers working in a decentralized and frequently offline workflow. Under that architecture, every local machine clones and stores the complete history of the project.
At Anthropic’s operational scale, that core assumption broke down. The main branch history was approaching 100 GB, alongside a compressed working tree exceeding 25 GB. With continuous integration (CI) pipelines running roughly 25 million checkouts every day, the company estimated it would have needed to transfer 600 petabytes of data daily without aggressive, multi-layered caching.
How Scion redesigned version control
To eliminate that overhead, Anthropic built an asymmetric client-server model that operates strictly online. Developer workstations and ephemeral agent containers never clone the full repository.
Heavy operations like commit history, diffs, line attribution (blame), and merges execute directly on the server through remote procedure calls (RPC). The command-line client (scn) fetches only the specific files needed through on-demand virtual checkouts.
Instead of Git’s packfiles and global repository locks, Scion tracks changes inside an append-only write-ahead log. Changeset metadata is indexed in Google Cloud Spanner, while compressed file blobs are stored in cloud object storage.
Claude models in the build process
A team of seven engineers developed and deployed Scion in five months. During that same window, the team built Antique, an internal code review platform, alongside a custom code search engine.
Anthropic relied on early research previews of internal Claude models to speed up development. In one case cited by the team, an unreleased preview of Claude Mythos rewrote the data ingestion pipeline that imported Git history into Scion. That single rewrite cut RAM usage by twenty times and ran four times faster.
The development workflow changed as well. Instead of queuing commits behind global repository locks, automated tests now run after code merges and before deployment. When simultaneous edits trigger merge conflicts, Claude steps in directly to reconcile the files.
Costs, practical limits, and closed code
Internal data from Anthropic indicates that Scion requires 200 times less CPU and memory than the fleet of Git mirrors the company previously ran. Checkout times dropped from minutes to seconds, freeing up 25% of overall CPU capacity across CI clusters.
Anthropic confirmed that it has no plans to open-source Scion or release it as a commercial product, noting that the software is tightly coupled to its proprietary cloud infrastructure. The migration highlights a structural shift for engineering organizations scaling up autonomous coders: once agents write software continuously, the bottleneck moves from typing code to the version control and validation systems that have to absorb it.