MCP SearXNG

What Ihor-Sokoliuk/MCP-SearXNG does

Ihor-Sokoliuk/MCP-SearXNG exposes SearXNG search and web-reading functions as MCP tools. It does not install or operate SearXNG itself; the MCP client must connect through an existing SearXNG service with JSON search enabled. The service may be self-hosted, operated by the user, or selected from a public deployment.

Search responses contain result titles, URLs, and content snippets. Agents can request later pages, restrict searches by language or time range, and select engines when the configured SearXNG instance supports the requested combination. A separate URL reader retrieves readable content from individual result pages, making the server suitable for workflows that first find sources and then inspect them.

How it works

The request path is an MCP client, the MCP server, SearXNG, and then the search engines configured by SearXNG. In local deployments, the client starts the server as a Node.js process or Docker container over STDIO. HTTP deployments expose the MCP service separately and use the full /mcp endpoint in the client configuration.

SEARXNG_URL identifies the SearXNG service, not the MCP endpoint. Multiple interchangeable SearXNG replicas can be configured as a semicolon-separated list. URL reading fetches the selected website directly rather than routing page content through SearXNG.

Privacy depends on the selected SearXNG deployment. An operator-controlled instance can reduce reliance on a third-party search operator, while a public instance receives the search query and may log it. Neither SearXNG nor this MCP layer independently guarantees anonymity.

Setup and configuration

Node.js 22 or later is required for NPX and npm installations. A minimal MCP client configuration starts the package with NPX and supplies the SearXNG base URL:

{
  "mcpServers": {
    "searxng": {
      "command": "npx",
      "args": ["-y", "mcp-searxng"],
      "env": {
        "SEARXNG_URL": "https://search.example.com"
      }
    }
  }
}

The package can also be installed globally with npm, run from the published Docker image, or built locally from the repository. Docker deployments must forward SEARXNG_URL into the container. Leave MCP_HTTP_PORT unset for local STDIO; configure an HTTP service when the client should connect to an independently running server.

Optional configuration supports replica failover or fan-out, caching, proxies, HTML fallback, and browser solvers. Docker deployments can pass FLARESOLVERR_URL and BYPARR_URL when those services are available. Browser-solver behavior uses FlareSolverr first and Byparr after eligible busy or transient-unavailable conditions.

Tools and capabilities

• searxng_web_search requires a parameter named exactly query. It returns search results with titles, URLs, and snippets, and accepts pagination plus filtering such as time_range and language.
• searxng_search_suggestions provides autocomplete suggestions for incomplete or unclear queries.
• searxng_instance_info reads reachable SearXNG /config endpoints to report categories, engines, defaults, locales, and plugins.
• web_url_read converts HTML to Markdown, pretty-prints JSON, and returns readable fenced text for plain text, YAML, TOML, and XML. It can return a whole page, content under a named heading, or headings only. Large pages can be paginated with startChar and maxLength.

PDF text extraction is bounded by input, output, page, time, concurrency, and memory limits. OCR is not supported. Binary, media, archive, and other octet-stream downloads are rejected rather than returned as raw data. Unreachable URLs or failed extraction produce an error string.