Muesli-HQ/muesli

Muesli – On‑device speech‑to‑text for macOS

What it is

• A native macOS application that provides dictation and meeting‑recording + transcription. It runs most of the speech‑to‑text pipeline locally on Apple Silicon (Apple Neural Engine) and falls back to hosted models (OpenAI, Anthropic, OpenRouter, Ollama, etc.) only when you opt‑in.
• The app also bundles a small command‑line tool (muesli‑cli) so other programs or AI agents can query recordings, transcriptions, and meeting notes.

Key capabilities

Feature	What you can do
Dictation	Hold a hot‑key (or double‑tap) to speak and have the text inserted where your cursor is. Latency ~0.13 s using the on‑device Parakeet model on the Neural Engine.
Meeting transcription	Record both your microphone and system audio, get VAD‑driven chunking, speaker diarisation, and a finished transcript in seconds. Summaries and titles can be generated with OpenAI, Anthropic (Claude), OpenRouter, ChatGPT, or local Ollama models.
Quill	Voice‑driven rewrite or generation of text at the cursor. Works with local or hosted LLMs.
Computer‑Use planner	Experimental voice‑controlled automation that runs local actions (open apps, type, click) using an on‑device Gemma model.
Live captions	Choose Apple Speech, Nemotron 3.5 or Parakeet Realtime for live‑preview captions during a meeting.
Model management	Download, delete and switch between many on‑device ASR models (Parakeet, Whisper, SenseVoice, Bodhan, Gemma, etc.) directly from the UI.
iCloud sync	Text, notes and metadata sync across macOS and iPhone; audio never leaves the device.
Shortcuts & Siri	Six pre‑made Shortcuts (start/stop dictation, start/stop meeting, fetch last dictation/notes) that can be triggered from Spotlight, Siri or keyboard.
CLI for agents	muesli‑cli lets scripts or LLM agents list recordings, transcribe files, fetch or update meeting notes, and output JSON/Markdown.
Privacy‑first	All on‑device processing by default; hosted services are optional and stream only audio, never store it.

Typical workflow

1. Install via the DMG or Homebrew (brew install --cask muesli).
2. On first launch, pick a default on‑device model (Parakeet Unified is suggested). The wizard checks microphone, accessibility and iCloud permissions.
3. Press the configured hot‑key, speak, release – the text appears where you typed.
4. For meetings, start recording from the UI or a Shortcut; the app captures mic + system audio, diarises speakers, and produces a transcript.
5. Optionally click Summarise to generate meeting notes with your chosen LLM (OpenAI, Claude, Ollama, etc.).
6. Sync the notes to iPhone via iCloud or export as PDF/Markdown.

Who it’s for

• Professionals who need fast, accurate dictation without sending voice data to the cloud.
• Teams that want automatic meeting minutes while keeping recordings private.
• Developers building AI agents that need a reliable macOS speech‑to‑text backend (via the bundled CLI).

Installation & building

• Pre‑built DMG or Homebrew cask for easy install.
• Source can be built with Xcode 26.6 (Swift 6.3) on macOS 14.2+. Build scripts handle the on‑device VQE runtime and app signing.

Limitations

• macOS 14.2+ required; many advanced live‑caption models need macOS 26+.
• On‑device models are sizable (hundreds of MB to several GB) and must be downloaded.
• The “Computer Use” planner is experimental and works only on macOS 15+ with Gemma models.

---

All details are taken directly from the repository’s README.