There is no embedding provider to configure and no key to obtain, which makes this cheap to try on documents you would not upload anywhere — the trade is that retrieval is similarity over a local index rather than a hosted vector service. The threshold argument on the query is what you tune when answers come back too loose or too empty.
A TypeScript MCP server that indexes documents on your own machine and answers questions against them by similarity. It handles PDF, DOCX, TXT and HTML, and is written with Chinese-language documents in mind as well as English.
- Ingestion — `add_document` for one file by path, `add_directory` to take everything in a folder
- `query_knowledge_base` — ask a question and get matching passages, with arguments for how many results to return and a similarity threshold to cut weak matches
- Management — `list_documents`, `get_document` for one by id, `remove_document`, `clear_knowledge_base`
- `get_stats` for what the index currently holds
- Automatic indexing on ingest, with the documents and index kept in their own directories beside the server
Node and npm — clone, `npm install`, `npm run build`, then point your client at dist/index.js. No API keys and no external service: indexing and retrieval both run locally. The server creates a documents directory and an index directory on first use, so run it somewhere you are happy for those to live. MIT licensed.
Build from source — clone the repository and build it, then point your client at the binary