Anthropic Scion replaces Git at 30,000 merges a day

Anthropic went from about 2,500 merged code changes a day in May to more than 30,000 in September. Git did not survive the trip. On October 13 the company published how it built Anthropic Scion, an in-house source control system that replaced Git in its monorepo. Scion is not for sale and not open source. The reason it exists is the part every software team should read: when your committers are agents, the tools built for humans become the bottleneck.

What actually happened

Per Anthropic's engineering post, a team of seven built Scion, a code review tool called Antique, code search and related systems, and finished the migration in five months.

The numbers that broke Git:

• The main branch history is almost 100 GB. A compressed working copy is over 25 GB.
• CI alone runs about 25 million checkouts a day. Anthropic says naive fetching would mean over 600 PB of Git traffic daily.
• At peak, the company ran several hundred internal Git mirrors to keep up.

The core mismatch: Git assumes offline, distributed clients. Almost all of Anthropic's clients are always-online agents and CI jobs. So Scion moves the work to the server. History, blame, commits, merges and rebases happen server-side, exposed through an API. A local working copy is optional and sparse.

Anthropic reports a CPU and memory footprint 200x smaller than its old mirror fleet, checkouts that went from minutes to seconds, and a 25% CPU capacity win in CI. It says Scion is "by no means a suitable replacement for git for general use." Instead, it wants to work with GitHub, GitLab and the open source community on server-side RPCs and better sparse and shallow features in Git.

Why agent-scale source control matters for your business

You do not have a 100 GB monorepo. But the curve is the same shape. Every coding agent you add opens more branches, triggers more CI runs, and pulls more full clones than a human ever would. The bill shows up in CI minutes, runner storage and slow pipelines, not in a line item called "agents."

What we would do now:

• Measure the clone. Check how long a fresh CI checkout takes and how often it runs. Shallow clones (--depth 1) and sparse checkouts are free wins.
• Cache dependencies, not just builds. Agents re-run the same jobs. Make the second run cheap.
• Put agent work behind the API. For small edits, agents can often use your forge's API instead of a full clone.
• Stay on standard Git. Anthropic's own advice is to wait for these ideas to land in GitHub and GitLab. Do not build a custom VCS.

Sources: Anthropic Engineering.