26,099 MCP Servers — and the Tiny Club That Actually Scores 90+

26,099 Servers, and Most of Them Don't Matter

The MCP catalog just crossed 26,099 servers. That number is the story everyone tells — the ecosystem is exploding, MCP won, etc. It's also the least useful number in this post.

Because volume isn't the signal. Quality is. And when you sort all 26,099 by the MCPpedia score — our 0–100 blend of security, maintenance, documentation, compatibility, and token efficiency — the top of the list is startlingly thin. Fewer than a dozen servers crack 94. Only two hit 95.

The 90+ Club

Here's who's actually at the top this week — every one of them active, zero CVEs, and a grade-A token footprint:

• Jscpd — 95. A copy/paste detector for 224+ languages, now Rust-powered. 1.15M weekly npm downloads. The highest-scoring server in the entire catalog.
• XcodeBuildMCP — 95. Xcode project + simulator control for agents. The iOS toolchain finally has a serious MCP story.
• Supabase MCP Server — 94. Databases, auth, storage, and edge functions from one server.
• Mcp Gitlab — 94. The GitLab API, cleanly wrapped.
• Brave Search — 94. Real-time search that doesn't leak your queries into an ad graph.
• PostgreSQL MCP Server — 94. The reference database server, still setting the bar.
• Tavily Mcp — 94. Production search, extract, map, and crawl in one place.
• Dbhub — 94. Zero-dependency, token-efficient access to Postgres, MySQL, SQL Server, and MariaDB.

Want to see how a 95 is actually built? Here's Jscpd's breakdown:

AI-ML Quietly Became the Biggest Category

For most of MCP's short life, developer-tools was the center of gravity. Not anymore.

AI-ML servers — agents wrapping other agents, model routers, memory layers, RAG plumbing — now outnumber developer tools by nearly 1,400. That's the ecosystem eating itself in the best way: MCP started as a bridge from models to tools, and it's increasingly a bridge from models to other models.

The category worth watching isn't the biggest, though. It's the smallest. Security sits at just 968 servers — under 4% of the catalog — even as agents get handed more credentials, more filesystem access, and more ability to act. The tooling to watch agents is badly outpaced by the tooling to empower them.

Stars Are Lying to You

Here's the most useful thing in the data this week. Sort by GitHub stars and you get a leaderboard dominated by projects that barely touch MCP: eslint-plugin-react-hooks (246K stars, score 55), n8n (194K stars, score 46). Huge repos, mediocre MCP servers.

Now sort by actual usage — weekly npm downloads — and a completely different cast shows up:

io.github.signerlabs/shipswift — 40+ production SwiftUI recipes delivered over MCP — pulls 16.5 million downloads a week with roughly 2,300 stars. Mcp_agent_mail, an async coordination layer that gives AI coding agents identities and shared inboxes, does 13.6 million — a hard signal that multi-agent workflows have left the demo stage and are running in real pipelines.

The Takeaway

The MCP ecosystem is not short on servers. It's short on good servers — and the gap between "published" and "worth installing" is widening every week.

So don't browse by count, and don't trust the star graph. Sort by score, check for a zero next to CVEs, and look at whether anyone's actually running it. Start with the 90+ club, read the scoring methodology so you know what the number means, and when you're stuck between two options, run them through the comparison tool.

26,099 servers. Your job is to find the eight that matter.