# Muesli

> Local-first AI dictation and meeting transcriber for macOS.

Muesli is a free, open-source macOS application that provides hold-to-talk dictation and meeting transcription powered by on-device AI inference. Speech-to-text runs locally on Apple Silicon by default, with optional hosted dictation providers. No cloud is required for core functionality.

## Features

- **Dictation**: Hold Left Cmd to speak, release to paste transcribed text at your cursor
- **Meeting Transcription**: Records mic + system audio, merges into a timestamped transcript
- **Meeting Notes**: Optional AI-generated structured summaries (BYOK OpenAI/OpenRouter)
- **Local-First**: Core transcription runs entirely on-device via Fluid Audio / CoreML
- **Privacy**: Audio stays local by default. Optional hosted dictation sends audio to the selected provider; hosted cleanup, Quill, and summaries send the input needed for those features when selected. iCloud sync transfers text and sync metadata, not audio
- **Free & Open Source**: MIT licensed, no subscriptions, no usage limits

## Technical Details

- **Runtime**: Native Swift/AppKit application (not Electron)
- **Speech Engine**: Fluid Audio SDK (CoreML, Apple Neural Engine)
- **Storage**: Local SQLite database at ~/Library/Application Support/Muesli/
- **Audio Capture**: AVAudioRecorder (mic), AudioHardwareCreateProcessTap (system audio)
- **Platform**: macOS 14.2+ on Apple Silicon (M1/M2/M3/M4)

## Links

- Website: https://freedspeech.xyz
- GitHub: https://github.com/Muesli-HQ/muesli
- Download: https://github.com/Muesli-HQ/muesli/releases/download/v0.8.4/Muesli-0.8.4.dmg
- Support: https://buymeacoffee.com/phequals7
- License: MIT

## Architecture

Muesli is a menu bar app with two modes:

1. **Dictation Mode**: User holds Left Cmd key → mic records → on release, audio is transcribed locally by default or by the selected hosted provider → text is pasted at the active cursor position via accessibility APIs.

2. **Meeting Mode**: User starts recording from menu bar → mic captures "You" audio, system audio captures "Others" → both are transcribed → segments are merged by timestamp → optional LLM summarization produces structured notes → everything stored in local SQLite.

The app uses Fluid Audio's CoreML models running on the Apple Neural Engine for speech-to-text, minimizing CPU/GPU usage. Optional hosted providers are available for dictation, cleanup, Quill, and meeting summaries when configured by the user.

## Author

Built by Pranav Hari (@pHequals7).
