Haypile: local document search for agents
Haypile is a single binary (hay) that indexes the user's folders into a local SQLite database and serves hybrid (semantic plus keyword) search on localhost:11500. The embedding model ships inside the binary. Outbound connections: 0.
Check it is available
hay status
The daemon autostarts on demand and listens only on localhost. If hay is not installed:
brew install BenyD/tap/hay
Windows (PowerShell): irm https://haypile.sh/install.ps1 | iex
Search, preferred path: MCP tools
If this client is connected to the Haypile MCP server, use its tools directly:
search_documents: hybrid search. Arguments: query (required), tag (optional, restricts to one indexed folder's tag), limit (optional). Returns ranked passages, each with its source file and page.list_sources: the indexed folders and their document counts.
Not connected yet? Add it:
claude mcp add --transport http haypile http://localhost:11500/mcp
Clients that launch a process instead of speaking HTTP can use the stdio bridge: command hay, args ["mcp-stdio"].
Search, fallback paths: CLI and REST
hay search "termination clause"
curl -X POST localhost:11500/api/query -d '{"query": "termination clause"}'
Index new content
hay add ~/path/to/folder # index a folder (or single file) and watch it for changes
hay list # what is indexed
hay remove ~/path/to/folder # un-index a folder
Indexing is local and incremental. A saved file is searchable within seconds. For a folder the user works in regularly, hay init writes a per-folder config (tag, exclude patterns) and can wire up .mcp.json; hay init --yes runs unattended.
Rules for answering from results
- Always cite. Carry each passage's file path and page into the answer.
- When exact wording matters (contracts, legal text, quotes), quote the passage rather than paraphrasing it.
- If search returns nothing, say so. Do not fill the gap from general knowledge. Offer to index the folder that would contain the answer.
- Scanned PDFs enter the index only when a local vision model handles OCR (
hay llm setup sets one up). If an expected document is missing from results, this is a likely cause; hay add reports skipped scans.
