Huddle records your meetings and turns them into transcripts, speaker-separated conversations, summaries, decisions and action items, entirely on your own Mac. No account, no cloud, no subscription. Put your MacBook on the table for an in-person meeting, or let it listen to the other side of a video call; press Stop and the notes write themselves while nothing leaves your computer.
- In the room or on a call: records the microphone and, optionally, the audio of every other app on your Mac (Teams, Zoom, Meet, FaceTime).
- Menu bar mode: close the window and Huddle keeps a small recorder in the menu bar. Click it or press ⌥⌘R to start and stop from anywhere.
- Crash-safe: the audio file is finalised every second, so a lost battery still leaves a recoverable recording.
- Import: drop in an existing audio file and get the same treatment.
- Transcript with speakers: who said what, with timestamps. Rename or merge speakers; Huddle suggests names it recognises from earlier meetings.
- Summary, topics, decisions: structured notes with evidence links back into the transcript, written in the language you choose (56 languages) whatever language was spoken.
- Action items: extracted on demand and tracked across meetings until you tick them off.
- Refine: tell Huddle what it got wrong and it rewrites the notes.
- Ask: ask one meeting or all of them a question.
- Everything runs locally: transcription, speaker separation and summaries all run on your Mac. Huddle works fully offline once the models are downloaded.
- Your data stays yours: recordings, transcripts and notes live in one folder under Application Support. Export any meeting as Markdown, text, SRT, JSON, or the original audio.
- Storage limit: set how much disk recordings may use; older audio is pruned first.
- MCP server: Claude Desktop, Claude Code, Cursor and other MCP clients can search your meetings, read transcripts and pull open action items, without touching your filesystem. Settings → MCP shows the ready-made configuration for each client.
- Network access (optional): share the MCP server on your local network with API keys.
- Bring your own models: Huddle finds models you already have (Ollama, LM Studio, Hugging Face cache) and only downloads what is missing.
Requirements: a Mac with Apple Silicon running macOS 14.2 or newer. About 1 GB of disk for the app plus the models you choose (the recommended set is around 5 GB). At least 16 GB of RAM is recommended, but small AI models will probably also work on 8 GB of RAM.
- Download
Huddle-<version>-macos-arm64.zipfrom the latest release, unzip it and dragHuddle.appinto Applications. - Open it. On first launch Huddle asks for the microphone and, if you want the other side of calls, system audio. It then checks what is already on your Mac and offers the model downloads it needs.
Requirements: macOS (Apple Silicon), Xcode Command Line Tools, Node 20+, Rust,
Homebrew Python 3.11 (python3.11).
# 1. engine
cd engine
python3.11 -m venv .venv
.venv/bin/pip install -r requirements.txt
.venv/bin/python -m pytest # tests
.venv/bin/python -m huddle_engine doctor # hardware, providers, models, resolution
# 2. system-audio helper (Swift; needs Xcode Command Line Tools)
../scripts/build-audio-tap.sh
# 3. speaker models (bundled into the app)
../scripts/fetch-speaker-models.sh
# 4. desktop app (starts Vite + Tauri; the app spawns engine/.venv automatically)
cd ../apps/desktop
npm install
npx tauri devHuddle stands on excellent open-source work:
- Transcription: Whisper via mlx-whisper and CTranslate2
- Speaker separation: sherpa-onnx with the pyannote segmentation model and NVIDIA's TitaNet embeddings
- Notes: Ollama running local models such as Qwen
- App: Tauri, React, Tailwind CSS, FastAPI
Huddle is created by me with extensive assistance from Claude Fable 5.1 (Anthropic) as a co-pilot for architecture, implementation and debugging.
Huddle is released under the MIT License.