An MCP server that helps AI agents find other MCP servers.

Ask in natural language — "what's available for sending Slack messages?" — and get back ranked servers with install commands and trust signals, instead of maintaining a hand-curated config file.

Ships with a bundled registry snapshot of 17,785 unique MCP servers (deduplicated from 18,366 raw records scraped 2026-02-09 from Glama.ai, npm, GitHub topics, and awesome lists). Works fully offline with zero API keys.

git clone https://github.com/yksanjo/mcp-discovery
cd mcp-discovery
npm ci
npm run build

Add to Claude Code:

claude mcp add mcp-discovery -- node /path/to/mcp-discovery/dist/index.js

Or to Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "mcp-discovery": {
      "command": "node",
      "args": ["/path/to/mcp-discovery/dist/index.js"]
    }
  }
}

Then ask your agent things like "find me an MCP server for querying Postgres" and it will call discover_mcp_server and get ranked, installable results.

In local mode, trust_score (0–100) is derived from GitHub stars and results carry no latency/uptime metrics — those require the hosted database. The two hosted-only tools fail fast with a clear message rather than crashing.

Honest caveats about the bundled data: install commands are best-effort from the scrape (verify before running anything), category labels are mostly missing, and the snapshot is a point in time — see "Refreshing the data" below.

For semantic search and live metrics you run your own backing services. There is currently no public hosted instance.

Required environment variables:

SUPABASE_URL=...                  # Postgres + pgvector (schema in src/db/)
SUPABASE_SERVICE_ROLE_KEY=...
OPENAI_API_KEY=...                # text-embedding queries

Setup: apply src/db/mcp-curator-schema.sql, then npm run seed and npm run generate-embeddings. DATABASE.md documents the schema. The REST API (npm run start, see src/api.ts and api/ for the Vercel entry) exposes the same search over HTTP.

• langchain/ — MCPDiscoveryTool for LangChain agents (async via aiohttp). Targets a hosted API endpoint you deploy yourself.
• AutoGPT blocks — MCPDiscoveryBlock / MCPDiscoveryCategoriesBlock.

The scrapers live in scripts/ (scrape_massive.py, scrape-glama.ts, and friends). They regenerate data_massive/mcp_servers_all.json, which local mode picks up automatically. You can also point MCP_DISCOVERY_DATA at any JSON file with the same record shape.

For allow/deny policies (e.g. excluding servers your org hasn't vetted), pass exclude_servers — see SECURITY.md.

npm run typecheck   # tsc --noEmit
npm run test:run    # vitest (includes an end-to-end local-mode discover)
npm run build       # compile to dist/

CI runs all three on every push and PR.

MIT