We Indexed 9,281 Claude Opus 4.7 Repos. Here’s What We Found.

Since Claude Opus 4.7 shipped, we’ve been harvesting every public GitHub repo it co-authors. As of today: 9,281 repos, 683K files, 121.7M lines of code. The clone rate is steady at ~400–500 repos/hr, so the corpus grows by ~10K/week.

This post is the executive summary. Follow-up posts drill into each axis.

The headline numbers

Metric Value
Repos indexed 9,281
Files analyzed 683,325
Total lines of code 121,787,946
Median repo size 7,722 LOC
Avg repo size 32,686 LOC
Unique primary languages 30+

What does Opus 4.7 actually build?

Based on LLM-parsed structured summaries of 413 analyzed repos:

Domain Share
Web backend 46%
Frontend 29%
CLI tool 12%
Data pipeline 6%
Game 5%
ML training 4%
Infrastructure 2%
MCP server (by domain) <1%

Opus 4.7 is a backend-leaning model. Despite its reputation for frontend work, nearly half of its output are servers/APIs. It also outputs CLI tools at a higher rate than almost any other AI coder we’ve seen.

The stack is TypeScript + Next.js + shadcn

Top five languages by repo primary language:
- TypeScript — 562 repos (+ 68,708 .tsx files)
- JSON — 783 (config/schemas)
- Python — 533
- Markdown — 526
- HTML — 370

Top actual imports (real library usage, not just declared deps):
- react — 45,938 imports
- lucide-react — 16,221
- @/lib/utils — 9,382 ← shadcn-ui signature
- next/server — 7,071
- @/components/ui/button — 6,828
- next/link — 6,739
- next/navigation — 5,926
- framer-motion — 3,790
- @tanstack/react-query — 3,418
- zod — 3,191

Opus 4.7 uses shadcn/ui by default. Across the corpus, ~19,000+ imports reference shadcn component paths. Lucide is the near-universal icon library. Next.js App Router dominates over Pages Router.

Quality profile

Against the community baseline (n=284K non-Opus repos):

Axis Opus 4.7 Community Delta
Overall score 60.2 58.1 +2.1
Code quality 79.8 79.8 0
Documentation 46.0 40.5 +5.5
Testing 27.3 21.3 +6.0
Practices 68.1 69.4 −1.3

Opus 4.7 writes noticeably better docs and tests. Raw code quality is indistinguishable from baseline — the ceiling is the detector, not the model. Practices are slightly behind (missing CI, hooks).

The Claude fingerprint

One signature stands out: Opus 4.7 repos contain CLAUDE.md or AGENTS.md at rates nothing else matches.

Fingerprint Repos
Has CLAUDE.md 1,324 (18%)
Has AGENTS.md 466
Has BOTH 371 (4%)
Name contains “mcp” 123
Name contains “agent” 102
Name contains “claude” 230

4% of Opus 4.7 repos are explicitly multi-agent setups (CLAUDE.md + AGENTS.md). This is not seen in Codex or community corpora.

Reuse potential

When the LLM grades a random Opus 4.7 repo for reuse potential:

  • HIGH — 72%
  • medium — 29%
  • low — 5%

Three out of four Opus 4.7 repos are considered “highly reusable.” That’s an order of magnitude denser training signal than generic GitHub dumps.

What’s next

Over the next few posts we’ll drill into:
- Why shadcn/ui is Opus 4.7’s default (and what it means)
- How Opus 4.7 writes functions (tl;dr: short)
- The CLAUDE.md phenomenon
- What’s missing: tests that run, dependency hygiene, CI
- The top 30 libraries Opus 4.7 actually imports

All insights refresh every 30 minutes as new repos arrive.