searxng-deepdive

An MCP server for SearXNG designed for LLM agents doing real research. Four tools with agent-friendly schemas, multi-page result fanout, lightweight URL→Markdown reading, and tool descriptions generated dynamically from the live engine pool of your SearXNG instance.

Why another mcp-searxng?

Existing packages are minimal — most expose a single search(query) tool with no way for the model to ask for more results, target specific engines, or constrain by category. The richer ones bake static descriptions, so the LLM never learns what's actually enabled on this instance. None of them treat agent-tool-selection ergonomics as a design priority.

searxng-deepdive opens those knobs up:

Quickstart

Install via npx -y from any MCP client:

{
  "mcpServers": {
    "searxng": {
      "command": "npx",
      "args": ["-y", "searxng-deepdive"],
      "env": { "SEARXNG_URL": "http://127.0.0.1:7979/" }
    }
  }
}

SEARXNG_URL should point at your running SearXNG instance. Need one? The companion repo SearXNG-Compose ships a plug-and-play Docker stack tuned for LLM consumption.

Tools

The server registers four tools. The LLM picks among them based on the descriptions below, augmented at startup with the live engine and category list from your instance.

search(query, [...])

Broad web search across the full enabled engine pool. Use when you don't have a specific source preference. Returns merged, deduplicated results across however many engines respond.

search_on_engines(query, engines, [...])

Search using only the specified engines (e.g. ["arxiv", "pubmed", "semantic scholar"]). The tool description registered with the MCP client includes the actual list of engines enabled on your instance — agents don't have to guess names. Validation rejects invalid names with a "did you mean" hint when they look like categories instead of engines.

search_by_category(query, categories, [...])

Search within specific categories — runs every engine tagged with each. Description includes the live category list and which engines belong to each. Same validation: invalid category names produce a clear error that points at search_on_engines when the offending value is actually an engine name.

web_url_read(url, [readHeadings, section, paragraphRange, startChar, maxLength])

Fetch a URL and convert its HTML to clean Markdown. Lightweight HTTP + HTML→Markdown (no headless browser) — handles ~80% of the static-HTML web (Wikipedia, docs sites, blogs, news, GitHub READMEs).

Token-efficient extraction modes (priority order, first set wins):

• readHeadings: true — return only the heading list as a hierarchical TOC
• section: "Installation" — return content under matching heading
• `paragraphRange: "3-

... View full README on GitHub