muesli

---

What is Muesli?

Muesli is a native macOS app for dictation and meeting transcription, built for Apple Silicon. Speak into any text field, capture a meeting without a bot, and turn transcripts into notes or a plan for your next task.

Speech recognition runs on your Mac by default. Hosted transcription and AI features are optional and send the content needed for the selected task to your provider. iCloud sync transfers text and sync metadata, never audio. See Muesli's privacy policy.

Features

• Dictation anywhere — Hold a hotkey, double-tap for hands-free recording, or use Press to toggle. Text is pasted at your cursor, with optional filler removal and cleanup.
• Failed dictation recovery — When dictation captures audio but transcription or buffered streaming finalization fails, keep the recording and any available partial text locally for playback and retry from Dictations. Retry uses your selected local model and saves text to history. Failed recordings remain until recovered or discarded; successful dictations keep no recovery audio.
• Microphone capture health — Standard dictation, Quill, and voice-command capture show an advisory warning after 3 seconds while startup remains pending. Core Audio events report confirmed capture failures; slow startup and silent audio alone do not fail a session. Startup telemetry contains only fixed status labels and timing, never recording content or device identifiers.
• Custom vocabulary and dictionary — Add names, technical terms, phrase matches, and replacement pairs to improve recognition. Import and export your dictionary as JSON for reuse in the app and CLI.
• Meeting transcription — Capture microphone and system audio with speaker labels. Enable live captions, import MP3/MP4/M4A/WAV files, or re-transcribe saved meetings with retained audio.
• Meeting video recording — Record a selected window or display with meeting audio and play it back locally. Requires macOS Screen Recording and Microphone permissions; obtain permission from meeting participants before recording. See meeting screen recording.
• Vidshot (experimental) — Record a selected window or screen region with microphone narration, then play the saved or recovered capture locally. Enable it in Experimental settings. See Vidshot.
• Meeting notes — Use templates, re-summarize with another model, and export notes, transcripts, and answers as Markdown or PDF.
• Ask Meetings — Chat with your meeting notes and transcripts directly beneath a saved meeting or from the Ask Meetings workspace. Use @ to attach saved meetings, continue with follow-up questions, and open cited passages to check the sources.
• Your choice of AI — Generate summaries with OpenAI or Anthropic API keys, OpenRouter, ChatGPT sign-in, Claude Code, Ollama, LM Studio, or a custom endpoint. Hosted providers receive the transcript for the requested task.
• Multilingual speech — Choose among Parakeet, Nemotron, Whisper, Qwen3 ASR, SenseVoice, Cohere, Bodhan, and Apple Speech. Bodhan Flex offers Native, Mixed, and Romanized output for Indic languages and English.
• Calendar and meeting detection — Show upcoming events from macOS Calendar accounts, choose calendars and time windows, and join or record from reminders. Optional camera and microphone activity detection recognizes meeting apps.
• Shortcuts and Siri — Start or stop dictation and meetings, retrieve the last dictation or meeting notes, and connect these actions to your own automations.
• Text sync with iPhone — Sync transcripts, notes, summaries, and dictation text through your private iCloud account. Audio stays on the recording device.
• Native Mac interface — Follow progress at the notch or in a floating indicator, browse meeting and dictation history, and manage models, shortcuts, and appearance in one app.

Hotkey dictation works while a meeting is recording or paused, producing the usual pasted text or voice-note output according to the selected mode. The meeting's local microphone track is silenced from an admitted explicit dictation start attempt until stopping, cancellation, or startup failure completes, including attempts that capture no dictation audio. Meetings that are recording continue to capture system audio; paused meetings stay paused. Refused requests remain meeting audio. If models need preparation, the indicators explain progress or failure; after “Ready”, retry dictation to begin recording.

Experimental features

Opt in under Settings → Experimental; availability depends on your build and model setup.

• Time Turner — Search local screen history by text, app, and date, generate source-linked daily recaps with your selected meeting model, and retrieve OCR and screenshots through muesli-cli time-turner.
• Computer Use voice commands — Control apps, browsers, and supported muesli settings with on-device Gemma on macOS 15+ or a connected ChatGPT account; AI prompt settings remain manual-only.
• Quill voice writing — Rewrite selected text, create text at the cursor, or ask a question using a local or hosted model.
• Cotypist typeahead (experimental) — Enable Cotypist in Settings for on-device inline writing suggestions through the muesli input source; Tab accepts a suggestion and ⌥→ accepts one word. Off by default. Accepted suggestions appear in your local Timeline with the preceding text and app name.

---

Install

Download the latest .dmg from Releases, open it, and move Muesli to Applications. Or install with Homebrew:

brew install --cask muesli

Muesli requires an Apple Silicon Mac running macOS 14.2 or later. Individual models may require a newer macOS version. Speech models download when selected and can be managed in the Models tab.

Use Xcode 26.6 (Swift 6.3) on macOS 26, matching CI. The app deployment target remains macOS 14.2.

git clone https://github.com/Muesli-HQ/muesli.git
cd muesli
./scripts/build_localvqe.sh
MUESLI_SKIP_SIGN=1 ./scripts/dev-test.sh

This installs an unsigned MuesliDev.app with a separate bundle ID and app data directory. Signed packages require a complete LocalVQE runtime in addition to the committed model. See CONTRIBUTING.md for build, signing, cache, and test instructions.

---

ChatGPT and Codex plugin

Install the official Muesli plugin to extract decisions and action items or turn a meeting into an implementation plan.

• With local shell access on your Mac: The agent can use Muesli's bundled CLI to find meetings and dictations, read transcripts, transcribe audio, and save revised notes when requested.
• Without local access: Export a meeting as Markdown and attach it to your conversation. Installing the plugin alone does not expose your Mac's meeting database.

Transcripts shared with ChatGPT or a hosted Codex model become part of that assistant's context and may be processed by its provider. Read the plugin setup guide and example prompts for the full workflow.

Agent CLI

The CLI ships inside Muesli.app and works with Codex, Claude Code, and other agents with local shell access. The agent can analyze transcripts using its own model without a separate Muesli summary API key.

export PATH="/Applications/Muesli.app/Contents/MacOS:$PATH"
muesli-cli spec
muesli-cli meetings list --limit 10
muesli-cli meetings get 42 # Replace 42 with an ID from the list
muesli-cli transcribe interview.m4a --format json

All transcription models in the app are also available through transcribe --model, including apple-speech (macOS 26 and compatible hardware), cohere-transcribe, bodhan-core, bodhan-core-int8, bodhan-flex, bodhan-flex-int8, gemma4-e2b, and gemma4-e4b (macOS 15+). Use muesli-cli transcribe --help for the full list. Downloads reuse the app’s model caches; Apple manages its own language assets. Apple Speech uses the system language, Cohere defaults to English, and Bodhan defaults to Hindi. Gemma retains its experimental status.

Data commands return JSON; file transcription returns plain text unless another format is requested. The meetings update-notes command replaces formatted notes while preserving the raw transcript. Use spec to check commands supported by your installed version.

See the CLI quickstart, coding agent guide, and full CLI reference for write-back, dictionaries, output formats, and troubleshooting. The repository's agent skill provides reusable instructions.

Custom vocabulary and dictionary

Use the app's Dictionary tab to add custom words and replacement phrases, or import and export entries as JSON. Imports merge by match word, update changed definitions, and append new entries. The same file can be used for CLI transcription:

muesli-cli transcribe interview.m4a --dictionary dictionary.json

Each entry specifies a word, its replacement, and a matching threshold:

[
  { "word": "museli", "replacement": "muesli", "matching_threshold": 0.85 }
]

---

Models

Choose and download models in the Models tab. English defaults to Parakeet Unified; multilingual options include Parakeet v3 and the families below.

Model family	Options	Approximate download
Apple Speech	System-managed dictation and live meeting transcription; macOS 26+	Managed by macOS
Parakeet	Unified, v2, v3, Redux, and Realtime EOU	220–565 MB
Nemotron 3.5	Multilingual live and final transcripts	665 MB
Cohere Transcribe 2B	Multilingual transcription	3.8 GB
SenseVoice Small	Multilingual transcription	240 MB
Qwen3 ASR	Multilingual transcription	1.3 GB
Bodhan Core / Flex	Indic languages and English; Pro or Lite; macOS 15+	2.46 GB / 1.27 GB weights
Whisper	Tiny and Small in English or multilingual variants, Medium English, Large Turbo Multilingual	153 MB–1.5 GB
Gemma 4 E2B / E4B	On-device dictation, cleanup, Quill and Computer Use; macOS 15+	3.6 GB / 5.2 GB
Muesli Cleanup Lite	Qwen 3.5 0.8B with extended fine-tuning and app and screen context; MLX 8-bit	820 MB
Muesli Pro Cleanup	Fine-tuned LFM 2.5, Qwen 3.5 or Gemma 4 with app and screen context	731 MB–2.6 GB
Cotypist	Qwen 3.5 0.8B base (default), Qwen 3.5 2B base, or Gemma 4 E2B base; MLX 4-bit	424 MB / 1.06 GB / 3.6 GB

Download sizes are storage estimates, not RAM requirements. Gemma's optional MTP assistant adds about 190 MB per model family; E2B and Gemma Pro Cleanup share an assistant. Gemma cannot be selected for transcription and cleanup simultaneously.

Gemma uses MLX Swift LM for generation and a native Swift/MLX audio encoder. Multi-token prediction (MTP) is on by default when verified assistant weights are available. If those weights cannot be downloaded or loaded, the main model remains usable with ordinary decoding. Computer Use keeps its tool-call grammar and validation in either mode. The app no longer packages LiteRT.

Gemma models prepare before recording, with progress shown in the recording indicator. First use and long idle periods can still take longer; further latency optimization is ongoing. Cleanup Lite and Qwen Pro Cleanup use the same native MLX-Swift-LM integration with 8-bit weights, as do Cotypist's base models (4-bit, no new runtime). LFM and S1-mini retain LLM.swift's llama.cpp runtime.

Live transcription is off by default. Apple Speech and Nemotron provide live and final transcripts; Parakeet Realtime EOU provides an English preview alongside a separate final meeting model. Enable it under Settings → Meetings → Transcription.

For background on the local runtimes, see Apple Neural Engine speech-to-text on Mac and the speech-to-text glossary.

---

Permissions and calendars

Onboarding guides you through the macOS permissions needed for the features you use.

Permission	Purpose
Microphone	Record dictation and meetings
System Audio Recording	Capture audio from meeting apps
Accessibility	Paste transcribed text
Input Monitoring	Detect global hotkeys
Camera activity detection	Recognize webcam activation for optional meeting detection
Calendar (optional)	Show upcoming meetings and reminders

For calendars, add Google, Exchange, or other accounts in System Settings → Internet Accounts with Calendars enabled. Grant Muesli Calendar access, then choose calendars under Settings → Meetings → Calendars. Muesli uses macOS Calendar through EventKit; direct Google Calendar sign-in is not available. Hiding a calendar in Muesli does not delete its events.

See Help and troubleshooting for setup and permission issues.

---

Tech stack

Component	Technology
Native app	Swift, AppKit, SwiftUI
Local inference	CoreML / Apple Neural Engine, MLX Swift / MLX Swift LM, Metal, LLM.swift / llama.cpp, WhisperKit
Audio and speakers	CoreAudio process tap, ScreenCaptureKit fallback, Silero VAD, pyannote diarization
Storage and sync	SQLite in WAL mode; CloudKit private database for text sync
Calendar	Apple EventKit
Automation	App Intents, Shortcuts, Computer Use planner, post-meeting executable hooks
Export	Markdown and PDF

Contributing

See CONTRIBUTING.md for development builds and tests, AGENTS.md for agent workflows, and the SQLite database guide before changing persistence or sync fields. Please open an issue before submitting a large PR.

Contributions require a DCO sign-off and disclosure of material AI assistance, as described in the contribution guide.

For manual watchdog and failed-streaming recovery QA, see Dev C failure simulations. The one-shot controls are Debug-only and do not alter microphone permissions or CoreAudio.

---

Support

For help, use the support guide or open an issue. If Muesli saves you time, consider supporting development:

---

Acknowledgements

Screenpipe informed Time Turner’s local visual-memory architecture. Muesli implements capture, OCR, and media storage independently in Swift using Apple frameworks.

Muesli has been possible because of the generosity of companies such as:

• FluidAudio — CoreML speech models for Apple devices (Parakeet TDT, Qwen3 ASR, Silero VAD, speaker diarization)
• localai-org/LocalVQE — on-device acoustic echo cancellation for meeting transcription
• WhisperKit — Swift Whisper inference on CoreML/ANE
• Core Audio by Apple — system audio process taps
• ScreenCaptureKit by Apple — system audio fallback capture
• Parakeet Redux by Moondream, based on NVIDIA Parakeet v3, with the Core ML conversion by FluidInference — model weights under CC BY 4.0. Muesli uses the converted weights without further modification.
• Phonon-2.5 by Fermion Research, with the Muesli Core ML conversion — CC BY 4.0. Uses the GPU encoder on macOS 15+; original notices and conversion attribution are included.
• MedASR by Google, with the Muesli Core ML conversion — English medical dictation under the HAI-DEF terms; see setup and runtime details.
• NVIDIA Parakeet — FastConformer TDT speech recognition model
• Cohere Transcribe — 2B parameter autoregressive ASR (#1 Open ASR Leaderboard)
• Qwen3-ASR — Multilingual speech recognition (52 languages)
• Bodhan AI Core and Flex — multilingual Indic/English ASR; community Core and Flex CoreML/MLX conversions
• MLX Swift — native Apple-silicon inference for Gemma and the Bodhan Pro/Lite hybrid runtimes
• MLX Swift LM — Gemma language-model inference and speculative decoding in native Swift
• Hugging Face Swift Transformers — native Swift tokenization for the MLX models
• LLM.swift and llama.cpp — local GGUF inference for LFM and S1-mini
• Gemma 4 E2B and E4B — MLX Community conversions used for on-device audio and text tasks
• pyannote — Speaker diarization (via FluidAudio CoreML conversion)

---

License

MIT — free and open source.

---

Resources

• Apple Neural Engine speech-to-text on Mac — how Muesli uses Apple Silicon, CoreML, and local ASR for fast dictation.
• Local speech-to-text glossary — ASR, VAD, diarization, acoustic echo cancellation, Parakeet, Whisper, and Qwen3 ASR.
• Best dictation apps for Mac — a practical comparison of Mac dictation tools.
• Offline dictation for Mac — why local-first voice typing matters.
• Local meeting transcription for Mac — meeting notes without adding a bot.
• Use Muesli meeting notes in ChatGPT and Codex — install the plugin and turn transcripts into notes and implementation plans.
• Hinglish and Indic language meeting notes on Mac — capture and review multilingual conversations with Muesli.
• Hinglish dictation on Mac — Hindi-English voice typing with Muesli.
• Bodhan Indic speech-to-text on Mac — Bodhan Core and Flex for multilingual dictation and transcription.
• Dictate in Hindi, Tamil, and other Indian languages — set up Indic language dictation on your Mac.
• Running Bodhan on Apple Silicon with CoreML and MLX Swift — how Muesli runs Bodhan locally.
• Muesli website · Documentation · Release history
• Privacy policy · Terms of service

---

Star History