Scion Replaced Git
Claude's code flood broke Git. So Anthropic built Scion.
Remember
When agents write the code, infrastructure is the bottleneck.
Context
Anthropic's monorepo: ~100 GB of history, >25 GB compressed working copy, ~25 million CI checkouts a day. Daily merged changes grew from ~2,500 (May 2026) to 30,000+ (September 2026), driven by Claude's autonomous coding. Fetching that the usual Git way projects to 600+ PB of traffic a day; at peak they ran several hundred internal Git mirrors. Source: Anthropic Engineering, "Redesigning source control" (2026-10-13), https://www.anthropic.com/engineering/redesigning-source-control
Architecture
| Component | Role | Notes |
|---|---|---|
| scion-daemon | Client on every machine | Bounded local cache, working copy, watches file changes; files materialized on demand (sparse by design) |
| scion-cache | Regional cache tier | Same RPC protocol as the server; >99% hit rate |
| scion-server | Authoritative server | Commits, history, blame, merges, rebases; API-first |
| Cloud storage + database | Blobs and metadata | Content-addressed compressed blobs; no fully materialized repo anywhere |
| Changeset | Scion's commit | Only the changed files vs. the first parent, pointers to blobs; created server-side, instantly visible |
| Derived data | Indices | Blame, path history, ancestry skip lists; computed incrementally, eager or deferred |
Trade-offs
- Always online: no offline work.
- The whole dev lifecycle must move (review, search, CI, IAM).
- Few refs (main + releases) to avoid global state.
- Server-side merges + post-merge testing accept more conflicts and lean on Claude to resolve them.
Numbers worth knowing
- 2,500 → 30,000+ merges/day (May → Sep 2026); ~100 GB history; ~25 M CI checkouts/day; 600+ PB/day naive projection.
- >99% cache hit rate per tier.
- 7 engineers, ~5 months, also built Antique (code review), code search and IAM.
- Claude rewrote the ingestion pipeline: 20x less memory, 4x faster.
- 200x smaller footprint than the Git mirrors; CI checkouts seconds instead of minutes (two orders of magnitude); 25% CI CPU saved.
