Whisnap
For builders and agents

An open engine.
For speech.

An open-source speech-to-text engine for your Mac. Load a model, transcribe, get text back. Use it from a script, an agent or your own app.

Open by design.
Publication on hold.

Engine not published yet. The public repository is unavailable.

Engine license: Apache-2.0. Model weights keep their own licenses.

No standalone install command is available yet.

Models change.
Keep the interface.

Parakeet runs on the Neural Engine through CoreML and FluidAudio. Whisper runs on Metal through whisper.cpp. Language-specific fine-tunes belong here too, including Lithuanian Parakeet.

Adding architectures and making them fast on Mac is the core work. CoreML and Metal are in use; MLX is part of the optimization direction. Each model still needs a supported backend and format.

Explore the model catalog

A catalog.
Not a permission slip.

Verified means tested on a real Mac. Curated means picked by the Whisnap team. Both are badges, not a gate: community models, local paths and Hugging Face repositories can run with an unverified warning.

Conversion, backend support and weight licenses still apply. Entries awaiting conversion or license review stay visible with their limitations.

CLI for scripts.
MCP for agents.

The engine, CLI and MCP are free. The CLI lists and downloads models, transcribes files to text or JSON, and can keep a model resident for repeated calls. MCP exposes a transcription tool over stdio and retains local models between calls.

The CLI ships inside the Mac app: choose “Install command line tool”. Homebrew distribution follows later. The public engine package is not published yet.

Command line guide

Embed it.

Use the Rust crate, C ABI or Swift package. The Mac app embeds the engine in-process, keeping a warm model in memory. The same engine can sit inside your own interface.

Apache-2.0 covers the engine. Model weights keep their own licenses and attribution requirements.

Better with real runs

Your Mac adds to the picture.

Opt-in run reports help people choose models for their Macs: model, hardware, timings and success or crash outcome. No audio. No text. Reporting is off by default, and sending is explicit.

Community speed appears only when a cohort has at least 20 successful reports from 5 batches. Reviews are separate: sign in, one review per model, with admin approval before publication.

Find your next model

If you want a UI for these models, there’s Whisnap.

Hold fn, talk, release. Text lands where your cursor is.

Download Whisnap