DEV Community

Edison Flores
Edison Flores

Posted on

Every MCP server in our index now has a public page — 66,496 of them, all crawlable

npx -y some-mcp-server runs code from a public registry on your machine. Before you do that, the honest questions are: who publishes this, how old is it, does anyone actually use it, and is the name a typosquat of something popular? Finding those answers today means chasing the repo, the package page, and the awesome-list entry — three tabs for a decision you make in ten seconds.

So we finished something we should have finished months ago: every MCP server in the MarketNow index now has its own public page.

What's on a page

Each page renders server-side (no JS required) and carries:

  • Trust score with signal breakdown — registry age, weekly downloads, typosquat distance to popular names, injection markers. A 100/100 page tells you why it's 100.
  • Install-risk classification — the honest three-tier: green (git-clone based, no registry code execution), yellow (heuristic signals advise review: young package, low adoption, stale repo, name similarity), red (installing via npx/uvx executes arbitrary code from the public registry).
  • Sources — repo and package links, version, author, indexed date.
  • Structured data — schema.org SoftwareApplication + FAQ JSON-LD, so the pages are machine-readable, not just human-readable.
  • A correction path — no signup. If your server's page is wrong, or you'd rather not be listed, there's a form. This matters more than the score.

What the catalog-wide numbers say

Across all 66,496 pages: 50,632 red, 15,861 yellow, 3 green. That distribution is not a bug — npx/uvx installs execute registry code by design, so an npm/PyPI-sourced server is red by definition unless its install path is git-clone based. "Red" is not "malicious"; it's "verify the publisher before running this." The 3 greens are the git-clone based ones. If someone's directory shows you a wall of green badges for registry-installed servers, that's not rigor, that's marketing.

Every entry has a trust score (mean 65.8), and the top categories are Developer Tools (31,366), Version Control (7,434), AI/ML (7,119), Search (4,846), Finance (3,676).

What the pages deliberately don't claim

A page says indexed and classified — never verified or trusted. Discovery is not authority: a listing means we found it and ran our heuristics; it does not mean the operator stands behind the endpoint at the moment you call it. That claim belongs to an evidence layer, which is a different project of ours and stays separate on purpose.

We also diffed ourselves against the awesome-mcp-servers curated list: we cover 3,541 of its 3,908 linked servers (91%). The 367 we're missing — plus 27 official npm packages (@heroku/mcp-server, @storybook/mcp, @mapbox/mcp-server, @doist/todoist-mcp...) — are queued for the next ingestion run. Publishing the gap is the point: a directory that hides its coverage holes is asking you to trust it blindly, which is exactly what a directory shouldn't do.

Why this matters beyond humans

The sitemap angle is the quiet big one. Before today, only 2,000 of these pages were declared to search engines. Now all 66,496 are. That's 66k long-tail entry points — "argocd mcp server", "dokploy mcp", "todoist mcp install" — each landing on a page that answers with a risk classification instead of a marketing badge. And the same URLs serve AI agents: the pages are structured-data-first, so an agent can parse the trust signals without scraping prose.

Search for an MCP server before you install it. If it's indexed, there's a page. If the page is wrong, there's a form. If it's missing, it's in the queue — and the queue is public.

(MarketNow is our project — the registry and the search API are free. Previous writeup on how the indexing pipeline works and the 2,910 fake install counts we deleted: 8,535 → 66,496.)

Top comments (0)