# Omni

A native macOS app that indexes text, code, PDFs, images, audio and video into one vector space for semantic search across a Mac's entire local filesystem, running fully on-device with no network connection required.

Omni embeds every kind of local file, text, code, PDFs, images, audio and video, into a single vector space using the jina-embeddings-v5-omni model, ported to run in-process on the GPU through MLX-Swift rather than through Python or a server. Because every file type shares one embedding space, a single query can surface any kind of matching file, and the app keeps its index updated as files change.

It browses like Finder, downloads its search model and an optional OCR model for scans and photos on first launch, and after that nothing leaves the Mac. A Serving setting exposes a local HTTP API and an MCP endpoint on the loopback address, so other tools or agents can query the same index. Building it from source requires Xcode 26 with the Metal Toolchain, and a test suite checks numeric parity against Python reference fixtures.

Website: https://hanxiao.io/omni