Privacy
Your meetings stay on your machine
Harknotes is built around a single principle: your audio never leaves your device. Here's exactly what stays local and what doesn't.
What stays on your machine
- All audio — mic and system audio are captured and processed locally. Recordings live on your disk and are removed when you delete the meeting. Never transmitted.
- Transcripts and notes — stored in a local SQLite database on your machine. Export any meeting as Markdown to a folder you choose. No sync, no cloud backup.
- Whisper model weights — downloaded once to your machine. Transcription runs entirely offline.
- AI summaries — by default, Harknotes talks to a local model through any OpenAI-compatible endpoint (Ollama, llama.cpp, LM Studio). In that setup, nothing leaves your machine at all.
What leaves your machine
Nothing — unless you opt in to a cloud AI provider for summaries. If you add your own API key for a hosted provider (OpenRouter, OpenAI, Groq, …), the text transcript and your typed notes are sent directly from your machine to that provider. No audio, no metadata — and no middleman.
How whisper.cpp works
Harknotes bundles whisper.cpp — a C++ implementation of OpenAI's Whisper speech recognition model — as a compiled binary. While you record, Harknotes cuts the audio into small chunks and runs this binary as a child process on your machine, so the transcript builds live during the meeting.
No internet connection is required for transcription. The Whisper model weights (downloaded once on first run) run entirely on your CPU or GPU.
No bot, no visible recording
Unlike Otter.ai, Fireflies, or tl;dv, Harknotes does not join your meeting as a participant. It captures audio at the operating system level — the same way screen recorders work. Other participants cannot see any recording indicator.