DEV Community

Cover image for 8,535 66,496: how we indexed every MCP server on GitHub, npm and PyPI (and deleted 2,910 fake install counts)
Edison Flores
Edison Flores

Posted on

8,535 66,496: how we indexed every MCP server on GitHub, npm and PyPI (and deleted 2,910 fake install counts)

#ai #security #opensource #mcp

In late June our registry listed 8,535 MCP skills. This week catalog v5.5.0 went live with 66,496 indexed MCP servers and 130,845 tracked ecosystem-wide — but the part I want to write about is not the growth. It's the garbage we deleted on the way.

What the numbers actually mean

  • 130,845 tracked — every MCP server we could see across three sources
  • 66,496 indexed — searchable in the registry, each with an install-risk classification (red / yellow / green) and source links
  • Deep-scanned subset — top packages get the 29 Sentinel rules: static analysis, dependency risk, maintainability and evidence-based trust scoring

The three sources are GitHub, npm and PyPI. Deduplication is by repository identity — a GitHub repo, its npm package and its PyPI wheel collapse into one listing.

The pipeline

Indexing runs in layers:

  1. GitHub — code and repo search for MCP-declaring repositories, verified via API (stars, issues, last push — real numbers only)
  2. npm — registry API scan for MCP packages, with download stats straight from the registry
  3. PyPI — project metadata and release history

Multi-registry indexing v2 (the release that took the index from 9k to 23k) landed earlier this week. v5.5.0 consolidated everything, re-ran the certifications, and purged the stale data layer.

What we deleted (the honest part)

The original auto-discovery inherited bad data. Two things were wrong:

  • 2,910 invented install counts — legacy listing rows whose install numbers could not be traced to any registry. All deleted.
  • ~2x inflated GitHub stars — catalog stars came from an old scrape; live API verification showed roughly double the real numbers. Every GitHub-sourced listing is now API-verified before it enters the catalog.

There's a transparency page that shows exactly where every skill comes from:

https://www.marketnow.site/catalog

If your MCP server is listed and something is wrong, there's a submission form — no signup, no account:

https://www.marketnow.site/submit

Use it from any MCP client 

npx -y marketnow-mcp
Enter fullscreen mode Exit fullscreen mode

That gives you 8 tools. The ones people use most:

  • search — query the 66,496-server registry by keyword and category
  • install-risk — red / yellow / green classification with the reason. Red means installing runs registry code on every start (the npx/uvx pattern) — flagged for review-before-install, not a verdict against the project
  • tool fingerprinting — hash a server's tools/list surface (RFC 8785 JCS + SHA-256) and diff it later to detect tool-poisoning redefinitions — the OWASP MCP cheat sheet item people keep asking about
  • credential verification — the 12-stage UTA pipeline covering 8 credential formats (JWT, W3C VC, MCP Card, A2A, EAT-AI, ZTA, X.509, ATC v3)

The searchable registry and the search API are free.

We told 44 maintainers yesterday

We opened 44 issues on repos whose servers are indexed — github-mcp-server, fastmcp, mcp-toolbox, awslabs/mcp, notion-mcp-server, firecrawl-mcp-server and 38 more. Each issue tells the maintainer their listing exists and includes a badge they can paste into their README. No signup, nothing to buy.

If you maintain an MCP server and want the badge: https://www.marketnow.site/embed

What's next

The trust pipeline (UTA) is getting its own site — it deserves not to live under a marketplace domain. More on that soon.

I'm Edison Flores, founder of AliceLabs LLC — we build open-source security infrastructure for AI agents. This post is about my own project; the registry, the search API and the badge system are free.

Top comments (0)