Everything that determines answer quality — which documents are loaded, how they are chunked, which model reads them — is configured in Langflow, not here. That is the appeal: you can rebuild the retrieval pipeline visually and the MCP side does not change. It also means the server is only as useful as the flow behind it, and `API_ENDPOINT` is the one setting that decides which flow that is.
A thin MCP server that forwards a question to a Langflow flow and returns the answer. The retrieval, the documents and the model all live in your Langflow instance; this makes that flow reachable from any MCP client as a single tool.
- A question sent to your Langflow Document Q&A flow, with the flow's response returned to the model — `query_docs`
- The flow chosen by configuration rather than by argument: `API_ENDPOINT` names which flow answers, so pointing at a different set of documents is a config change, not a code change
A running Langflow instance with a flow built from its Document Q&A template — chat input, file upload and an LLM component — and that flow's API endpoint URL, which Langflow shows behind the API button as a cURL command. Put that URL in `API_ENDPOINT`; it defaults to a local Langflow address with a placeholder flow id, so it must be set to reach your own flow. The npm package name is `doc-qa-server` (0.1.0) and it is marked private, so build from a clone with `npm install` and `npm run build`, or use the Smithery installer.
One command — npx -y github:GongRzhe/Langflow-DOC-QA-SERVER