The GTM MCP
Directory

Ask for the job, not the category.

An agent does not want a data enrichment tool. It wants a person's title from a LinkedIn URL. Type the thing you are trying to do. This runs in your browser over a baked index: no backend, no query logging, and it keeps working with the network cable pulled out. The box and the chips write q, mcp, gate and cli onto this page's URL so a link opens the same filter. Those query strings are not sitemap pages. The contract is on the data page and in llms.txt.

Ordering is fixed and published, never tuned and never purchasable. official MCP first, then community, then unknown, then n/a, then none-found; within each band gate order is free, paid, enterprise-leaning, enterprise-only, unknown; then alphabetical by name. Computed, never curated, never purchasable. With a query typed, matches are banded by relevance first and the rule above breaks every tie. An exact name match pins to the top and nothing else is boosted. The filters above run over the 1,236 unique products, so their totals sit below the 1,252 entry totals used by the MCP status and gate views. Both numbers are correct and the difference is the 16 cross listings.

Point your agent at the directory.

The server loads the baked file once at import and answers from memory. It makes zero outbound network requests, so it cannot be slow, cannot rate limit you, cannot cost anything, and cannot leak your query to a vendor. Everything network shaped happens in the weekly build.

{
  "mcpServers": {
    "gtm-directory": {
      "url": "https://andrewcmcguire.com/gtm-directory/api/mcp",
      "headers": {
        "Authorization": "Bearer gtmd_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"
      }
    }
  }
}

That is the hosted copy, live since 2026-09-08: streamable HTTP, no install, the same read-only server this page is built from, restarted on every publish. It needs a free key, and gtmd_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx above is where yours goes. Request a key. A work email address is approved automatically, usually within about ten minutes. The hosted copy records, per key, the number of calls and the date last used, and nothing else: no query text, no tool arguments, no IP log kept.

claude mcp add --transport http gtm-directory https://andrewcmcguire.com/gtm-directory/api/mcp --header "Authorization: Bearer gtmd_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"

Same key, as one line for Claude Code. A client that only accepts a URL (claude.ai custom connectors) cannot send a header, so it uses the per-key URL instead:

{
  "mcpServers": {
    "gtm-directory": {
      "url": "https://andrewcmcguire.com/gtm-directory/api/mcp/k/gtmd_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"
    }
  }
}

Both forms also work on the fallback host https://d1hkopq5aq852m.cloudfront.net/gtm-directory/api/mcp. Some clients hit a Cloudflare 403 on the apex host today; if yours does, swap the host and keep the rest. If you would rather run it yourself, no key is needed, because the code and the data are public:

{
  "mcpServers": {
    "gtm-directory": {
      "command": "uvx",
      "args": [
        "gtm-mcp-directory"
      ]
    }
  }
}

The package is not on PyPI yet, so this block is the shape the install will take rather than a working one-liner today. The server source is real and public: it lives in the gtm-mcp-directory repo and runs from a checkout right now.

The hosted /api/mcp path has no backend LLM and no model selection. The model is always your MCP client's model, including a local model in LM Studio. How to use the hosted MCP has the prompts, the LM Studio steps, the honesty vocabulary, and the limits.

Answers, not opinions.

The questions people actually ask about GTM tools, MCP servers and agents, answered from this data with the numbers generated at build time and the date stamped on every one. No tool versus tool verdicts, because 1 tools here have been bench tested.