Omni mascot, a small 3D mole

Search every file on your Mac by meaning.

Text, code, PDFs, images, audio, and video - in any language, entirely on-device. Download the model once, then run fully offline.

Download for macOS v0.2.1

Apple silicon · macOS 14+ · MD5 33460c01b88c82a33cc8bda54cd56bf3

37 seconds: search by meaning, any file to any file, deep PDF search, folder maps, and the on-device architecture.

Features

Any language

Search in English, match notes in German, Chinese, or Japanese - queries and files meet by meaning, not by keyword, scanned pages included.

See a folder as a map

Every file becomes a dot, clustered by similarity and colored by type. Click a dot to spotlight its nearest neighbors - duplicates, drafts, and forgotten corners stand out at a glance.

Fast on Apple silicon

A native MLX-Swift port of jina-embeddings-v5-omni runs in-process on the GPU - tens of thousands of tokens per second on recent chips, measured in the open rather than claimed.

Native to macOS

List and Gallery views, real QuickLook thumbnails, drag and drop, keyboard-first navigation, and filters by kind, folder, date, and score.

Private by design

Everything stays on your Mac.

The model download is the only time Omni touches the network. From then on your files, their embeddings, and every query stay on this Mac - pull the plug and search still works.

One model download, then fully offline
No accounts, no telemetry
Local SQLite vector store you can inspect or delete
Open source on GitHub

A look inside

Search with any file, find any file.

Every file - image, audio, video, text - lands in one shared vector space, so a query is just another file. Drop in a photo to find related screenshots, a voice memo to surface where it came from, a PDF to pull up everything near it. Right-click any result and choose Find Similar, or drop a file onto the search. One model, one space, any direction.

Searching images - Porsche 911 photos — porsche 911 · images jina omni embeddings · text podcast interview · audio screen recording demo · video this file · results across every modality Documents · mapped by similarity

Searching text and code - jina omni embeddings — porsche 911 · images jina omni embeddings · text podcast interview · audio screen recording demo · video this file · results across every modality Documents · mapped by similarity

Settings - Content: file types to index, size thresholds, ignore rules — Choose what to index.

Settings - Performance: model choice, memory cap, folder map layout — Tune memory and speed.

Settings - Serving: the local HTTP embedding server — Run a local embedding server.

Settings - Indexing: live activity and throughput — Watch indexing live.

Settings - Storage: where the index lives and how big it is — One index, on your disk.

Settings - History: search history retention — History on your terms.

Serving

Give your local agents search.

Omni exposes its search over your indexed files as a local endpoint - loopback-only and token-guarded. Agents like Hermes and OpenClaw search the index you already built - on your machine, with no cloud round-trip.

Search endpoint

The same semantic search the app uses - how agents actually reach your files.

Embeddings · add-on

Raw vectors too, via OpenAI-, Jina-, Cohere-, and Gemini-compatible APIs.

Benchmark

How fast is Omni on your Mac?

Community benchmarks on a fixed 300-file dataset. One line per chip, one dot per Omni version - hover for details. Add your Mac: File → Run Profiling in the app.

Loading community benchmarks…

Questions & how-to

Everything Omni does, and how. Click a question to expand.

Loading…