Local Rag MCP Server

DuckDuckGo and Google lookups are available through rag_search_ddgs and rag_search_google, which accept a query and return Markdown context from selected web pages. Text embeddings help rank search results against the question before content is extracted. Callers can adjust the initial result count, choose how many ranked results to retain, and decide whether to include source URLs. The server runs locally without API keys while retrieving information from the web.

Broader investigations use deep_research to combine related search terms across chosen backends, such as Bing, Brave, Wikipedia, Google, and DuckDuckGo. It aggregates the searches, ranks content by relevance, and removes duplicates. The deep_research_google and deep_research_ddgs shortcuts keep multi-query research within their respective engines. Research calls also expose per-term retrieval and retention limits plus optional source URLs. Implemented in Python with stdio transport, mcp-local-rag has been tested with Claude Desktop, Cursor, and Goose. Its included Agent Skill guides Claude on query formulation, tool selection, and research parameters.