Your files, your computer, your AI.
Garage indexes your documents, code, notes and messages on your computer and serves them to your AI assistant over the Model Context Protocol. Garage uploads nothing: your assistant gets only the excerpts it searches for, never the whole collection.
~/Library/Application Support/GarageApp first.
What Garage does
Indexes what you already have
Folders, git repositories, Markdown, PDF, Office documents, scanned images and Apple Messages and Mail, parsed into searchable Chunk A passage of a document. Garage embeds and searches chunks, so search returns the part that matches, not the whole file. in a private PostgreSQL database.
Hybrid search
Vector similarity and full-text search fused with Reciprocal Rank Fusion A way to merge rankings — each result scores 1 ÷ (60 + its rank) in every list, and the scores are added up., so a question finds the meaning and a keyword finds the exact line.
Knows who wrote it
Git history, document metadata and path rules classify every document as authored by you, reference material or received from someone else, with the evidence recorded.
Works with your AI assistant
An MCP 2.0 server lets Claude Desktop, Claude Code, Cursor and other MCP The Model Context Protocol, the open standard AI assistants use to call tools such as Garage's search. clients search your Corpus Everything Garage has indexed from your sources — documents, code and messages — searchable as one collection., read documents and ask grounded questions.
Local models, your choice
Embeddings and fact distillation run on the built-in llama.cpp engine, or on the Ollama or LM Studio server you already have. Switch models without re-ingesting.
Private by construction
No cloud AI client in the app. One tested egress choke point with a destination allowlist, and Garage itself never sends your messages off your computer. Read the privacy policy →
A look inside
Search that finds the meaning and the exact line
Every result shows how it matched, whether it is yours or someone else's, and the full text beside it.
Answers grounded in your own files
Try a question in the app before your assistant does: a local model answers from what search returns, and cites it.
Connects to the assistants you already use
Garage finds the MCP clients on your computer and connects each one with a click.
Always knows where your library stands
Documents, chunks, embeddings and distilled facts, counted on the Status page.
Where your data goes
Garage never uploads your files or its index. What leaves your computer depends on which AI you connect, and Garage’s privacy guarantee is enforced by tests, not by promise.
Claude, ChatGPT and other cloud assistants
Most people connect a cloud assistant. Its app on your computer queries Garage over MCP and gets only the relevant parts of your collection: the excerpts its searches return and the documents it opens, never the whole index. Garage uploads nothing, but the assistant does send those parts to its provider's servers with your conversation, under the provider's terms, so leave out any source you don't want to go there.
Fully private, on your computer
Use the built-in llama.cpp engine on macOS, or LM Studio or Ollama on this computer, for embeddings and answers, and connect an agent that runs its model locally, such as LM Studio's chat. Nothing leaves the machine: not your files, not the index, not the questions you ask.
Private AI on your own network
Point Garage at an LM Studio or Ollama server on another machine you run, and documents and code go to that one server for embedding, and nowhere else. Messages and Mail never do: they stay on this computer even then.
How it works
Garage walks the sources you register, extracts text, decides who wrote each document and how much to trust it, and splits it into chunks. Each chunk is embedded under every model you register, one table per model, so adding a model is a Backfill Embedding the chunks you already have with a newly added model, without reading your files again. rather than a re-ingest. Search fuses the vector and keyword rankings and serves the result to your assistant over MCP. Everything lives in a PostgreSQL 18 + pgvector cluster the app bundles and runs for you.
Download
Garage for macOS
Apple Silicon (M1 and later), macOS 14 Sonoma or later. Free on the Mac App Store, or as a signed and notarized .pkg installer from this site.
Get Garage from the Mac App Store, or download the installer from this site. Both versions share one library on your computer, so you can switch between them without re-indexing. The App Store version gets its updates from the App Store instead of the in-app updater.
After installing
- Open Garage from Applications. It appears in the menu bar and starts its private database.
- The first-run assistant picks your folders, an embedding model and the AI clients to connect.
- Ask Claude, or any MCP client, a question about your own files.
The installer version checks for updates through Sparkle, only after asking you once; the App Store version updates through the App Store. The macOS version runs on Apple Silicon only. The .zip archive is on the GitHub release page.
Windows and Linux
Windows alpha
A Windows version is on the way. The database and the indexer already build and pass their tests on Windows; the first alpha will be posted here.
About the Windows alpha →Linux command line
There is no desktop app for Linux, but the whole pipeline runs there: the garage command, the garage-mcp server and hybrid search, over PostgreSQL with pgvector.
Also a command line
The macOS app includes a garage command for terminal workflows and a garage-mcp stdio server for MCP clients (in Garage.app/Contents/MacOS). The same pipeline builds as those two commands with Cargo, so the indexer, extractors and search run anywhere PostgreSQL with pgvector does, Linux included. The developer documentation covers the architecture, attribution engine, privacy internals and database schema.
Open source, on GitHub
Garage is free and open source. Everything is in one repository on GitHub: the macOS app, the Rust pipeline, the Windows app and this website.
Source and releases
Read the code, download every release and pre-release, and follow along as the Windows version comes together.
rickmark/garage-rag ↗Issues and ideas
Report a bug, ask a question or suggest a feature. The app's Report a Bug button fills in the details for you.
Open an issue ↗Contribute
Pull requests are welcome, from typo fixes to new extractors. Testing on Windows and Linux helps a lot right now.
How to contribute →The Support Center has the user guide, troubleshooting, the FAQ and how to report a bug.