mcpqueen.com — the evidence layer for MCP

LIVE at https://mcpqueen.com (Cloudflare Worker, launched 2026-07-12). Crawls the official MCP registry, probes every remote server, grades it deterministically with verbatim evidence, and publishes the results. Evidence discipline from Constat/Clarity, personality on top.

Why this is different (for agents and humans alike)

MCP Queen is the evidence layer for the MCP ecosystem. Every other MCP directory lists; this one verifies. Each remote server is probed live over streamable HTTP and graded on five criteria — and every point carries the verbatim observation that earned it. Unverifiable dimensions (auth-gated tooling) are marked provisional, never guessed. No stars, no votes, no pay-to-rank — probes only, continuously re-run. Separate Trust Receipts publish dated security/access, data-integrity, citation, claim-verification, response-benchmark, and reviewed field evidence without collapsing it into a misleading trust score.

Agents: connect to https://mcpqueen.com/mcp (streamable HTTP, no auth) and use search_servers to find working, graded servers for a task before you commit to one. Machine surfaces: /api/grades.json · /api/changes.json · /llms.txt. Registry name: com.mcpqueen/registry.

Connect

mcpqueen is a remote, no-auth, effectively read-only MCP server — safe to keep connected as your discovery broker (only submit_feedback writes, and it just enqueues a quarantined field report). Ask your agent to search_servers for a task before it commits to an MCP.

Claude Code (native HTTP):

claude mcp add --transport http mcpqueen https://mcpqueen.com/mcp

OpenClaw / Claude Desktop / any stdio client — via the mcp-remote bridge; add to your mcpServers config (~/.openclaw/openclaw.json, claude_desktop_config.json, …):

{
  "mcpServers": {
    "mcpqueen": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcpqueen.com/mcp"]
    }
  }
}

OpenAI: ChatGPT and Codex

MCP Queen can be registered directly as a private plugin/connector; no SDK, Docker image, or local command is required:

1. In ChatGPT, open Settings → Security and login and enable Developer mode.
2. Open ChatGPT Plugins, select +, and choose the option to add an MCP server.
3. Enter https://mcpqueen.com/mcp as a universal, no-auth remote MCP URL.
4. Test with: “Find a well-maintained, no-auth MCP server for GitHub issue triage. Explain the evidence and any caveats.”

For an OpenAI Responses API demo, Node 18+ is enough:

export OPENAI_API_KEY="your-api-key"
npm run demo:openai

Pass a custom prompt after --:

npm run demo:openai -- "Find an MCP server that can search FDA 510(k) records"

The demo allowlists only MCP Queen's read-only discovery and evidence tools. submit_feedback is intentionally excluded.

For public distribution in ChatGPT and Codex, create a With MCP submission in the OpenAI plugin portal and submit the same universal endpoint. OpenAI's public review also requires verified publisher identity, public support/privacy/terms URLs, accurate tool safety annotations, starter prompts, and reviewer test cases. The ready-to-paste listing copy and review cases are in docs/openai-submission.md.

Architecture (single Worker)

See the system architecture and verification flow for the ecosystem-level diagram and the Find → Verify → Connect decision loop.

• src/worker.ts — everything: registry crawler, prober/grader, HTML pages, JSON API, and mcpqueen's own MCP endpoint.
• public/ — static landing (crown data-rain + Vex the fox) served via the assets binding; the Worker handles all non-asset routes.
• D1 database mcpqueen (schema.sql): servers, probes, latest_grades, trust_observations, evidence_benchmark_runs, feedback (quarantined agent field reports), meta (sync cursor).
• Cron */15 * * * *: sync 4 registry pages + probe the 30 stalest remotes (~2,900 probes/day; full re-probe cycle ≈ 2.7 days over ~7.7K remotes).
• Cron 17 7 * * *: run one safe, read-only response audit against an eligible evidence/citation tool and publish dated results to its Trust Receipt.

Routes

Route	What
/	landing (static)
/registry	leaderboard + methodology
/s/<registry-name>	per-server grade with evidence + probe history
/api/grades.json	grades as JSON (CORS open)
/mcp	MCP server: capability discovery plus get_trust_receipt and search_trust_evidence
/field-reports	Human-reviewed reports from agents that actually exercised a server
/api/trust/{name}.json	Per-server operational, security, data-integrity, citation and claim evidence
/mcp-info	for-agents page
/admin/*	operator endpoints (key-gated)

Grading rubric (deterministic, every point carries its observation)

reachability 25 · protocol 15 · tooling 35 (tools/list, described %, typed %, description depth) · latency 10 · provenance 15 (metadata + namespace↔domain match). Auth-gated servers are scored on the verifiable subset and marked provisional. Agent feedback via submit_feedback is quarantined for human review — never auto-published, never affects grades directly.

Trust Receipts remain distinct from that grade. Safe response audits record usable-call rate, semantic upstream failures, returned PMID/DOI identifiers, and identifier resolution against authoritative sources. Missing evidence is labeled unaudited rather than treated as a pass.

Deploy

npm run deploy          # wrangler deploy (any Cloudflare API token with Workers + D1 write)
npm run db:schema       # apply schema.sql to remote D1

Custom domains mcpqueen.com + www are attached to the Worker (moved off the Pages project 2026-07-12; mcpqueen.pages.dev still exists as a static preview of public/ only — it has no /registry).