What it does
Haypile watches the folders you point it at, indexes their documents and returns ranked passages through MCP or REST, each with a source file and page citation. Search combines meaning-based and keyword matching, so paraphrases and exact identifiers both match. It ships as one binary with the embedding model included, and it also works as a standalone command-line tool. Text documents are indexed and searched without a network connection.
It suits people whose agent, such as Claude Code or Cursor, needs to answer from local files without reading whole documents or uploading them to a cloud service. The hay init command sets up a per-folder configuration, indexes the folder and can write an .mcp.json file so the agent can search it. Answers from hay ask are optional and are generated from the retrieved passages.
Running it means installing the binary and, for scanned PDFs, a local vision model for OCR. The daemon starts on hay add and listens only on localhost. Answers need either a local OpenAI-compatible server, such as Ollama, LM Studio, llama.cpp or Jan, or a cloud endpoint you choose with your own key. Without an answer model, hay ask shows the top passages instead.
Limitations and requirements
Scanned PDFs need a local vision model for OCR. Generating answers needs a local OpenAI-compatible server or your own cloud API key; cloud answers send only the passages retrieved for that question. API keys are refused over plain http to any host other than localhost. Without the embedding model, which development builds load from disk, search runs in keyword-only mode.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
Haypile
Hybrid search over your local documents, with file and page citations. One binary, fully local.
v0.3.1As shown on registry.modelcontextprotocol.io when we last read it