Launching today
Garage
Your files, your Mac, your AI. Private RAG over MCP
1 follower
Your files, your Mac, your AI. Private RAG over MCP
1 follower
Garage is a free, open-source Mac app that turns the files you already have into a searchable library for your AI assistant. It indexes folders, git repositories, PDFs, Office documents, scanned images, Apple Mail and Messages into a private database on your Mac, and keyword search over the Model Context Protocol. Claude Desktop, Claude Code, Cursor, LM Studio and any other MCP client can connect with one click. Garage uploads nothing: your assistant receives only the excerpts it searches for.
I built Garage because I wanted Claude to answer from my own notes, code and email without uploading any of it.
It's a Mac app that indexes folders, git repos, PDFs, Office files, scanned images, Mail and Messages into a Postgres + pgvector database it runs itself, and serves hybrid search (vector + full text, fused with Reciprocal Rank Fusion) over MCP. Any MCP client can connect: Claude Desktop, Claude Code, Cursor, LM Studio.
Some things that were fun to build:
Postgres running inside the App Store sandbox. That needed a flock() data-directory interlock and pthread semaphores in place of System V IPC; I've prepared them as a patch series for upstream Postgres.
One egress choke point: a single module may import a network client, an AST test enforces it, and communications can never be sent to a host that isn't loopback.
Authorship attribution from git history, document metadata and mail senders, so search knows what you wrote versus what you received.
It's MIT-licensed, Apple silicon only, and free (there's an optional tip jar in the App Store build). I'd love feedback, especially on retrieval quality and on which file types you'd want next.