convez.ai
MCP ServerIntegração & APIsIntermediário
yksanjo

MCP Discovery

yksanjo/mcp-discovery

Descoberta natural de servidores MCP com índice offline de 17k+

Ver no GitHub
BomPluginScore · de 0 a 100

Servidor MCP que ajuda agentes de IA a descobrir outros servidores MCP através de buscas em linguagem natural. Funciona offline com um índice bundled de 17k servidores, e oferece modo hospedado com busca semântica e métricas ao vivo.

Instalação

$npx convez-ai yksanjo/mcp-discovery

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
8
Forks
3
Acompanhando
0
Saúde
85%
Último commit
há 5 meses
Versão
v1.2.0
Issues abertas
0
Contribuidores
3

Casos de uso

  • Encontrar servidores MCP para tarefas específicas sem manutenção manual
  • Integrar buscas semânticas em agentes IA que precisam de ferramentas externas
  • Comparar múltiplos servidores MCP por métricas de confiabilidade

Veredito

Solução pragmática para o problema real de descoberta em um ecossistema fragmentado de MCP servers. O modo offline funciona imediatamente; o modo hospedado requer infraestrutura própria, mas a separação clara entre recursos locais e remotos é bem pensada.

README

A tradução para o português deste README ainda está na fila.

mcp-discovery

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

CI License Last Commit

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.

Quick start (local mode, no 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.

Tools

ToolLocal modeHosted mode
discover_mcp_server✅ keyword search over the bundled snapshot✅ semantic search (OpenAI embeddings + pgvector)
get_server_metrics❌ (needs live probe data)✅ latency / uptime / success-rate history
compare_servers❌ (needs live probe data)✅ side-by-side rankings

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.

Hosted mode (optional)

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.

Integrations

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

Refreshing the data

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.

Development

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.

License

MIT

Ver o original no GitHub

Bom para

Relacionados