Offline meeting notes — record and transcribe on-device

Daisy makes offline meeting notes: the recording, the transcript, the speaker labels, and the search all run on your own machine, and the audio and transcripts never leave it. No bot joins your call and nothing is uploaded. The AI layer can stay local too — point Daisy at a model on your own hardware and even summaries never leave your disk.

What runs on your machine

These steps are fully on-device — you can record and transcribe with the wifi off:

• Recording. Daisy captures system and mic audio straight to disk. No calendar, no meeting link, no bot in the room — you press record. Open formats: .wav for the source and .opus for a compact playback copy.
• Transcription. Speech-to-text runs on-device. You get a timestamped markdown transcript when the meeting ends.
• Speaker ID / diarization. Figuring out who spoke when is always on-device and always free. It never leaves your machine and never costs a credit.
• Keyword search. Once a transcript exists, searching across your notes is local — grep-speed over your own files.

That covers the whole core loop: capture, transcribe, separate speakers, find it later. None of it sends your audio or transcript anywhere.

What needs a model (and how to keep that local too)

There is exactly one thing that wants a language model: the higher-level stuff — summaries, Q&A over the transcript, and coaching-style feedback. A raw transcript doesn't need a model; turning it into a summary does.

You have two honest paths, and one of them keeps the AI fully local:

• Run a local LLM. Point Daisy at Ollama or LM Studio on your own machine and summaries, Q&A, and coaching are 100% local. The transcript and the model both live on your disk, so the AI layer sends nothing out either.
• Bring a cloud key (BYOK). If you'd rather use OpenAI, Anthropic, or Groq, Daisy sends only the transcript text — over HTTPS, with your own key, no proxy in the middle. That path uses the internet, by definition, because the model lives somewhere else.
• Use no LLM at all. Daisy hands you the transcript plus a ready-made, copy-paste prompt. Take it to whatever tool you like — or nothing at all.

This is what "private by default, your choice to extend" means: the on-device core is the default, and the cloud is something you opt into deliberately, key in hand.

Your notes live on your disk

Everything Daisy produces is a plain file in a folder you chose. Audio is .wav and .opus; transcripts and summaries are markdown. There's no proprietary database to export from and no server-side copy to delete. Back them up, grep them, drop them in your own vault, or check them into a repo — they're yours, in formats that will still open in a decade.

No bot in your call, no per-seat meter, and your recordings never reach us.

Works when you're offline

The core loop doesn't need a connection, so being offline is fine:

• On a plane. Record an interview or a working session at 35,000 feet with the wifi off, and walk off with a finished, searchable transcript.
• Spotty or no connection. Recording, transcription, speaker labels, and search keep working while you're disconnected; a local LLM keeps summaries working too.
• No-cloud-for-your-data policies. When sending audio or transcripts to a third party is off the table — compliance, contracts, classification — Daisy keeps that data on the device. (The AI stays local only if you run a local model; a BYOK cloud key sends transcript text out by design.)

One license covers three devices, on Linux, Windows, and macOS (Apple Silicon .dmg), with a 30-day trial.