MCP Server Registry — Health & Trust Scores

A catalog of Model Context Protocol servers that answers the question every list skips: before I wire this into an agent — is it maintained, what transport does it speak, how do I install it, and should I trust it?

Built on the free public GitHub API. No login, no anti-bot.

Why this one

MCP went from a proposal to tens of thousands of servers in about a year. The registries that exist give you a name and a star count. Neither tells you that the server you're about to depend on was last touched 14 months ago, has no license, and speaks a transport your client doesn't support. This is the "npm registry data" of the agent era: a catalog with a maintenance verdict.

What you get

Field	Description
trustScore	0–100 "would I ship this?" — deliberately weights maintenance over popularity
maintenanceHealth	0–100 from push recency, with small bonuses for a license and real adoption
abandonmentRisk	active · slowing · at_risk · likely_abandoned · archived
daysSincePush	How stale the code actually is
transport	stdio · sse · http (or unknown) — what your client must support
installMethod	npx / npm · uvx / pip · docker · go install · cargo install
fullName / url / owner	Repository identity
stars / forks / openIssues	Adoption signals
language / license / topics / description	Metadata
archived	Dead by definition (excluded unless you ask for them)

Why a 40-star server can outrank a 5,000-star one: a popular server abandoned two years ago is a worse dependency than a small one shipped last week. trustScore reflects that on purpose.

How to use it

1. Add a free GitHub token. Unauthenticated search is rate-limited after a page or two; a token raises the limit ~30×. Read-only public scope is enough.
2. Leave topics empty to sweep the standard MCP topics, or pass your own.
3. Tick Only actively maintained (and/or set a minimum trust score) to get a shortlist you'd actually ship.

Input

{
  "githubToken": "ghp_your_token",
  "activeOnly": true,
  "minTrustScore": 50,
  "minStars": 5,
  "maxPages": 2
}

Sample output

[
  {
    "fullName": "acme/mcp-thing",
    "trustScore": 84,
    "maintenanceHealth": 100,
    "abandonmentRisk": "active",
    "daysSincePush": 5,
    "transport": "stdio",
    "installMethod": "uvx / pip",
    "language": "Python",
    "stars": 120,
    "forks": 8,
    "license": "MIT",
    "description": "MCP server over stdio",
    "topics": "mcp-server, python",
    "url": "https://github.com/acme/mcp-thing",
    "archived": false,
    "status": "ok"
  }
]

Typical uses

• Pick a dependency — shortlist servers by trust before adding one to an agent stack.
• Audit what you already use — feed your current servers' topics and check none have gone stale.
• Ecosystem research — transport mix, language mix, and how fast the space is actually maintained.
• Monitoring — schedule it; a server sliding from active to at_risk is an early warning.

Pricing

$1.50 / 1,000 servers ($0.0015 per result), plus a near-zero start fee. Never charged beyond maxItems; rate-limit and search problems are reported in QUALITY_REPORT, not billed as data.

FAQ & limitations

• Do I need a GitHub token? Not strictly, but without one GitHub will rate-limit you quickly. It's free.
• Where does discovery come from? GitHub topics. A server that doesn't tag itself with an MCP topic won't be found — that's a real limitation, and it's why the actor reports how many raw hits it saw.
• Is transport guaranteed? No — it's inferred from the description/topics and can read unknown. Treat it as a strong hint, and confirm in the server's README before wiring it up.
• Does it test that servers actually run? No. It scores maintenance and metadata, not live uptime. Nothing here executes third-party code.
• Why are archived repos missing? Excluded by default; tick includeArchived to see them.

Related actors

• Reliability Toolkit (MCP) — your own MCP server, exposing site-diligence tools to agents.
• GitHub Repo Scraper · npm Package Info — the same health approach for general dependencies.
• AI Crawler Audit — the other half of the AI-readiness picture.