RAGnarok: Web Search MCP for Open-Source LLMs
Overview
RAGnarok is a web search MCP server packaged as a single Docker container. It uses SearXNG for search and Playwright for page retrieval, then returns structured results over MCP stdio — bringing real-time, self-hosted web access to open-source LLMs with no proprietary search API keys and no closed-source dependencies. Drop it into any MCP-aware client (LM Studio, OpenCode) and a local model becomes a web-connected assistant.
Tools
- web_search(query, top_k) — searches the web through SearXNG, scrapes allowed pages, and returns ranked results with title, source, content, published date, and score.
- read_source(url, max_pages) — reads a public URL directly and optionally expands into a bounded set of nearby same-origin pages, returning extracted content.
Responsible by design
RAGnarok scrapes conservatively: only public HTTP/HTTPS targets, respects robots.txt, never clicks consent prompts, TOS dialogs, or sign-in gates, and filters blocked and interstitial pages out of results. The whole retrieval pipeline is deterministic — the same query against the same pages returns the same chunks in the same order, which is exactly the property you want when wiring search into an autonomous agent loop.