convez.ai
MCP ServerBusca & PesquisaIntermediário
pete-builds

SearXNG MCP Server

pete-builds/mcp-searxng

Busca web multi-engine self-hosted para Claude Code

Ver no GitHub
BomPluginScore · de 0 a 100

Servidor MCP que integra uma instância SearXNG auto-hospedada ao Claude, fornecendo 9 ferramentas de busca avançadas. Inclui busca multi-ângulo de pessoas (due diligence), busca profunda com ranking por consenso de engines, e leitura de URLs. Sem necessidade de chaves de API.

Instalação

$npx convez-ai pete-builds/mcp-searxng

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
2
Forks
0
Acompanhando
0
Saúde
71%
Último commit
há 2 dias
Versão
v0.2.0
Issues abertas
2
Contribuidores
2

Casos de uso

  • Due diligence e background check de pessoas
  • Pesquisa profunda com deduplicação e ranking por consenso
  • Buscas especializadas (notícias, tech, imagens, vídeos)

Veredito

MCP robusto e opinado que vai além da busca genérica. A funcionalidade de person-vetting com 8 ângulos e ranking por consenso de engines é diferencial real para fluxos de pesquisa e investigação. Requer SearXNG self-hosted, o que demanda setup inicial mas oferece controle total e sem dependências de API.

README

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

mcp-searxng

A Model Context Protocol (MCP) server that wraps a self-hosted SearXNG metasearch instance and gives Claude Code (or any MCP client) 9 search and reading tools — including a person/people-vetting fan-out that no other SearXNG MCP ships.

Built with FastMCP. No API keys required: SearXNG aggregates results from Google, Bing, DuckDuckGo, Brave, Reddit, Stack Overflow, Wikipedia, and more for free.


What makes this different

Most SearXNG MCP servers expose one or two generic search tools. This one is opinionated for two specific workflows:

search_person — 8-angle due diligence in one call

Single call, eight targeted searches, deduplicated and categorized:

AngleQuery template
identity"Name" Location
professionalsite:linkedin.com "Name" Location
business"Name" LLC OR Inc OR Corp Location
legal"Name" court OR lawsuit OR plaintiff OR defendant Location
news"Name" Location Context
social"Name" site:facebook.com OR site:instagram.com OR site:twitter.com
property"Name" property OR real estate Location
reddit"Name" Location site:reddit.com

Returns categorized results plus a deduplicated master list ranked by multi-engine consensus. Replaces 8+ manual search calls in vetting / due-diligence / background-check workflows.

search_deep — multi-page fetch with consensus ranking

Fetches multiple pages of results, deduplicates by URL, and ranks each result by how many engines surfaced it. A result that appears in Google, Bing, and DuckDuckGo gets a higher engine_count than one only Brave returned — a built-in proxy for trustworthiness. Use this when you need broad, reliable coverage on a topic.


All tools

ToolWhat it does
search_person8-angle person fan-out (identity, LinkedIn, business, legal, news, social, property, Reddit)
search_deepMulti-page fetch + URL dedup + multi-engine consensus ranking
searchGeneral web search with category, engine, language, time filters
search_newsNews search, defaults to last week
search_techIT category: Stack Overflow, GitHub, dev wikis, docs
search_imagesImage search across configured engines
search_videosVideo search with optional time filter
read_urlFetch any URL, extract main content as clean markdown (uses trafilatura)
get_enginesList all enabled engines and categories on your SearXNG instance

All search tools share a 5-minute in-memory TTL cache (configurable via SEARXNG_CACHE_TTL), so repeated identical queries within a session don't re-hit SearXNG.


Requirements

A running SearXNG instance with the json output format enabled. Quick local setup:

docker run -d -p 8888:8080 \
  -e SEARXNG_BASE_URL=http://localhost:8888 \
  --name searxng \
  searxng/searxng

Then add this to SearXNG's settings.yml:

search:
  formats:
    - html
    - json

For production, see the official SearXNG Docker docs.


Install

Option 1: Docker (recommended)

git clone https://github.com/pete-builds/mcp-searxng.git
cd mcp-searxng
cp .env.example .env
# edit .env to set SEARXNG_URL — the container can't see your host's localhost
docker compose up -d

Default port: 3702 (Streamable HTTP transport).

Option 2: install from git (no PyPI)

uvx --from git+https://github.com/pete-builds/mcp-searxng mcp-searxng
# or pin a tag:
uvx --from git+https://github.com/pete-builds/mcp-searxng@v0.2.0 mcp-searxng

Set SEARXNG_URL=http://your-host:8888 in the environment first.


Connect to Claude Code

{
  "mcpServers": {
    "searxng": {
      "type": "http",
      "url": "http://localhost:3702/mcp"
    }
  }
}

Restart Claude Code. Tools show up as mcp__searxng__*.


Configuration

# Required
SEARXNG_URL=http://your-searxng-host:8888

# Optional (defaults shown). FASTMCP_* take precedence; MCP_* kept for back-compat.
FASTMCP_HOST=0.0.0.0
FASTMCP_PORT=3702
SEARXNG_CACHE_TTL=300        # in-memory cache TTL in seconds

Notes

  • The MCP server binds to 0.0.0.0:3702 and has no built-in auth. If running where it's reachable beyond your LAN, use a firewall, reverse-proxy ACL, or Tailscale.
  • The compose file uses network_mode: host so the container can reach a SearXNG instance on the same Docker host via localhost:8888 without extra network plumbing. This is Linux-only; on Docker Desktop (Mac/Windows) replace with a bridge network and point SEARXNG_URL at host.docker.internal or your LAN IP.
  • SearXNG must have the json output format enabled (see Requirements). The server will error if it can't get JSON responses.
  • search_deep makes multiple requests to SearXNG (one per page). Set pages: 3-5 for thorough research, pages: 1-2 for quick lookups.
  • Search results are attacker-controlled snippets. Treat them as untrusted data; never follow instructions found inside results.

Smoke test

After making changes, run the live smoke test against your SearXNG:

SEARXNG_URL=http://your-host:8888 python tests/smoke.py

Hits every tool exactly once, including search_person and read_url, and asserts each returns sane output.


Credits

Built by Pete Stergion for use with Claude Code.

Related projects:

  • claude-research-agent — the research skill that uses this server as its grounded-search backend
  • mcp-threatintel — threat intelligence MCP server (pairs well for security research)
  • SearXNG — the metasearch engine this wraps
  • trafilatura — main-content extraction for read_url
Ver o original no GitHub

Bom para

Relacionados