mcp-local-rag is a local MCP server implementing a primitive RAG-like web search context protocol. It runs locally with no APIs exposed, designed to integrate with MCP clients that support tool calls. When an LLM submits a query, the server uses DuckDuckGo to fetch search results (10 results per query), obtains embeddings from Google's MediaPipe Text Embedder, and computes similarity to rank the results. It then selects the top results and extracts context from the URLs by converting HTML content to Markdown for the LLM. The resulting contextual content is returned to the language model, which can generate a final response that incorporates fresh web information. The repository provides a Mermaid diagram illustrating the end-to-end flow, and deployment options via uvx or Docker. Docker deployment uses the container image ghcr.io/nkapila6/mcp-local-rag:latest. Overall, this is a lightweight, API-free MCP server intended for local experimentation with retrieval-augmented generation.