Muesli

Muesli is a local-first audio recorder, transcription app, and batch transcription tool. It is inspired by Granola's workflow, but it is not a full Granola clone.

It can:

• record from your microphone
• transcribe speech locally with Whisper
• summarize transcripts with a local GGUF model or an optional API fallback
• process existing audio files in batch
• write sidecar .txt and .silent files for downstream tooling

The current repo is Windows-first, but the core Python code also runs on macOS and Linux with the right audio and model dependencies.

Why This Repo Exists

Muesli started as a local meeting recorder and evolved into a practical transcription tool for poor-quality real-world audio, shared audio folders, and sidecar-based workflows.

The repo currently includes:

• a desktop GUI recorder: muesli_gui.py
• a reusable Python API: muesli.py
• a batch transcriber for 2026-* style audio drops: muesli_batch_transcribe.py
• a Windows global hotkey listener: muesli_hotkey.py

Current Status

The project is usable, but it is still a working tool rather than a polished packaged product.

• Windows setup is the best-supported path right now.
• The batch transcription workflow is production-useful.
• The GUI and launcher path are functional, but still evolving.
• There is no installer package or release build in this repo yet.
• The app intentionally focuses on local recording, local transcription, and shared-folder workflows instead of trying to match Granola feature-for-feature.

Granola Comparison

This project borrows some UX and workflow ideas from Granola, but it is narrower in scope. Granola's current product includes calendar-aware meeting capture, templates, recipes, workspaces, integrations, and mobile sync. Muesli currently focuses on local recording, local transcription, local summaries, and file-based batch processing.

The table below is based on Granola's public docs for transcription, calendar sync, templates, recipes, workspaces, integrations, and iPhone sync.

Feature	Granola	Muesli	Notes
Local desktop recorder	Yes	Yes	Both have a desktop capture workflow.
Live transcript view	Yes	Partial	Muesli shows live transcript while recording, but its UX is simpler.
Mic + system audio capture separation	Yes	No	Granola distinguishes your mic from system audio on desktop. Muesli is mic-first and batch-file oriented today.
Calendar-aware upcoming meetings	Yes	No	Granola syncs Google and Outlook calendars. Muesli does not.
Quick ad-hoc notes	Yes	Yes	Granola has Quick Note; Muesli can start a manual recording immediately.
AI-enhanced notes / summaries	Yes	Yes	Muesli can summarize locally with GGUF or via optional API fallback, but Granola's note-generation workflow is more advanced.
Template-based note regeneration	Yes	No	Granola supports desktop templates and re-generation. Muesli currently uses a single editable prompt.
Saved prompt recipes / chat over notes	Yes	No	Granola exposes recipes and cross-note chat. Muesli does not yet.
Shared workspaces and folders	Yes	No	Muesli currently uses filesystem conventions rather than multi-user workspaces.
Web sharing and collaboration	Yes	No	Granola supports shared links and folder collaboration.
Integrations (Slack, Notion, CRM, Zapier, MCP, API)	Yes	No	Muesli currently integrates through files and scripts rather than productized SaaS integrations.
iPhone capture and desktop sync	Yes	No	Muesli is desktop-only today.
Local batch transcription from shared folders	No	Yes	This is one of Muesli's strongest workflows.
Sidecar transcript and no-speech markers	No	Yes	Muesli writes .txt and .silent files for race-free downstream processing.
Fully local audio retention	Partial	Yes	Granola states it uses a transcription provider and does not save audio. Muesli keeps audio and transcripts locally under your control.

Features

• Local recording at 16 kHz mono
• Local transcription with faster-whisper
• Optional local summarization with llama-cpp-python
• Optional Anthropic fallback if you configure an API key yourself
• Batch processing of existing audio files
• Sidecar output convention:
  • slug.txt for detected speech
  • slug.silent for processed audio with no useful speech
• Windows global hotkey support for launch-and-record

Quick Start

Windows

git clone https://github.com/joshwhitk/Muesli.git
cd Muesli
setup_windows.bat

That script:

• creates .venv
• installs Python dependencies
• checks for ffmpeg
• creates desktop and startup shortcuts

After setup, launch the app with the desktop shortcut or run:

.venv\Scripts\python.exe muesli_gui.py

macOS

Requires Homebrew and Python 3.10+.

git clone https://github.com/joshwhitk/Muesli.git
cd Muesli
bash setup_macos.sh

That script installs ffmpeg and portaudio via Homebrew, creates a .venv, and installs Python dependencies.

After setup, launch the app:

source .venv/bin/activate
python muesli_gui.py

macOS notes:

• The OS will prompt for microphone permission on the first recording.
• Transcription runs on CPU (the fast Whisper model is recommended for speed).
• The global hotkey agent is Windows-only; use Ctrl+R inside the app to start/stop recording.

Linux

Linux is manual setup at the moment:

git clone https://github.com/joshwhitk/Muesli.git
cd Muesli
python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt

You will usually also want:

sudo apt install ffmpeg portaudio19-dev

Then run:

python muesli_gui.py

Batch Transcription Workflow

The batch transcriber is designed for shared audio directories and poor-quality recordings.

Typical convention:

2026-04-15_18-21-46.wav    recording
2026-04-15_18-21-46.txt    transcript with speech
2026-04-15_18-21-46.silent processed, no useful speech

The local mirror under outputs/sidecars/ is intentionally ignored by Git.

Configuration

Runtime configuration lives in config.json and is not committed.

Common fields:

{
  "shared_dir": "C:\\Users\\<you>\\Documents\\MuesliData\\analytics\\audio",
  "launch_hotkey": "Ctrl+Shift+`",
  "whisper_quality": "high",
  "whisper_model": "large-v3",
  "whisper_device": "auto"
}

Notes:

• shared_dir controls where exported audio and sidecar files go.
• launch_hotkey is used by the Windows hotkey listener.
• whisper_quality is the user-facing preset shown in Settings.
• whisper_model and whisper_device are the underlying Whisper runtime choices.

Models

Whisper

Whisper is provided by faster-whisper. The model is downloaded and cached automatically on first use.

• High Quality uses large-v3
• Fast uses medium
• New installs default to High Quality when there is at least 20 GB of free disk space on the local user drive
• On Windows with an NVIDIA GPU, setup_windows.bat also installs the CUDA runtime packages needed for GPU Whisper

Local summarization model

If you want local summarization, place a GGUF model in models/.

Example options:

• Phi-3.1-mini Q4
• Mistral 7B Instruct Q4
• another small instruction-tuned GGUF that works with llama-cpp-python

If no local summary model is available, the code can fall back to Anthropic if you provide an API key in config.json.

Ollama on Windows

When Muesli uses Ollama for summaries, the GPU-heavy work runs inside the global Ollama service (ollama.exe / ollama.exe runner), not inside the Muesli process itself.

• Task Manager will usually show that work as ollama.exe, not as Muesli
• Muesli cannot reliably rename those processes or force Task Manager to group them under Muesli without changing the execution model
• The practical fix is better in-app visibility: show the active backend, model, and runtime status clearly inside Muesli

Windows Hotkey

Muesli ships with a small background listener on Windows so the launch-and-record shortcut works system-wide.

Current default:

Ctrl+Shift+`

This is handled by RegisterHotKey, so it works across normal desktop apps and is not limited to Explorer shortcut hotkeys.

Development

Basic setup

python -m venv .venv
.venv\Scripts\python.exe -m pip install --upgrade pip
.venv\Scripts\python.exe -m pip install -r requirements.txt

Smoke tests

There is a basic smoke test script:

.venv\Scripts\python.exe test_muesli.py

It is closer to an integration smoke test than a unit test suite.

Syntax validation

This repo includes a lightweight GitHub Actions workflow that compiles the Python files to catch syntax errors on push and pull request.

Repo Layout

muesli.py                   core API
muesli_gui.py               desktop GUI
muesli_batch_transcribe.py  batch transcription tool
muesli_hotkey.py            Windows global hotkey listener
setup_windows.bat           Windows setup helper
MUESLI_API.md               API notes and examples
assets/                     icons and branding assets
models/                     local GGUF models (not committed)
recordings/                 local metadata/audio working files (not committed)
outputs/                    batch outputs and sidecar mirror (not committed)

Known Gaps

• No packaged installer or signed release build yet
• No formal migration path for older Granola-era paths/configs
• Tests are still smoke-test oriented
• The GUI still has some Windows-specific rough edges

Task List

• Build a proper Windows installer
• Add a settings panel for model and runtime configuration
• Improve packaging so the pinned taskbar icon no longer depends on Python
• Expand tests beyond smoke checks
• Add export to Obsidian workflow
• Add a simple transcription progress diagram in the session view
• Fix speaker counting for multi-speaker clips such as "I guess I'm going to"
• Add clearer Ollama runtime visibility in-app so users can tell when background GPU usage comes from Muesli-driven summaries
• Show a blinking red recording dot over the Muesli logo in visible recording-state icons, starting with the tray icon and then desktop/taskbar assets where practical

Documentation

• API notes: MUESLI_API.md
• Contribution guide: CONTRIBUTING.md

License

MIT. See LICENSE.