Labsco
MCP SERVER

MCP Searxng Relay

by littleoffice ↗

Web search and page fetching through your own SearXNG instance, with a per-caller ledger of which URLs were actually fetched and how much of each was read.

Web Search Engines
Summary
It records what the agent read, not just what it searched.

The source list distinguishes a page read in full from one where only metadata was fetched — the distinction models lose first when they compose a citation far from the tool call that produced it. Long documents are windowed at 100k characters with the exact next `start_index` returned, and the extracted text stays cached, so paging through a large PDF costs one upstream fetch rather than one per page. Reaching an internal wiki is opt-in through `FETCH_ALLOWED_HOSTS` or `FETCH_ALLOWED_CIDRS`, and both require an explicit port.

What it is

A search relay for environments where queries are not allowed to leave infrastructure you control. It fronts a self-hosted SearXNG instance, fetches and extracts pages itself, and keeps a per-caller record of what it retrieved — so a claim about sources can be checked against a list rather than trusted to a model's recall.

What you get
  • Web search through your SearXNG instance with control over language, category, time range, safe-search level and result count, and the backend engines that surfaced each result named on it — `searxng_web_search`
  • A URL fetched and converted to structured Markdown with headings, lists, tables and code blocks preserved; PDFs and Office documents (DOCX, XLSX, PPTX plus legacy DOC, XLS, PPT) are extracted too, and images come back as image content for vision models — `searxng_read_url`
  • Title, author, publish date, language, site name, description, image, categories and tags for a URL as JSON, for choosing between candidate links before paying for a full read — `searxng_url_metadata`
  • The URLs this relay genuinely fetched for the caller, newest first, each marked with how much was read: full text, one window of a longer document, metadata only, or a failed fetch — `searxng_session_sources`
Requirements

A running SearXNG instance with the JSON output format enabled in its `settings.yml`, and its base URL in `SEARXNG_URL` — the server refuses to start without it. Docker (`ghcr.io/littleoffice/mcp-searxng-relay:latest`) or Go 1.26+ to build from source. With `MCP_PORT` set it speaks Streamable HTTP and requires a bearer token via `MCP_AUTH_TOKEN`, `MCP_AUTH_TOKENS` or `MCP_AUTH_TOKEN_FILE`; with `MCP_PORT` unset it runs over stdio.

Setup effort

One command — docker run -d \ -e SEARXNG_URL=https://your-searxng-instance.example.com \ -e MCP_PORT=8080 \ -e MCP_AUTH_TOKEN=$(openssl rand -hex 32) \ -p 8080:8080 \ ghcr.io/littleoffice/mcp-searxng-relay:latest