Local AI meeting notes, on your Mac

Most AI meeting tools record on your computer and then send the audio to a server for transcription and summaries. That is fine for many meetings. It is not fine for a client meeting under legal privilege, a therapy session, an interview with a source, or a board meeting about something that has not been announced.

For those, you want the work done where the recording is: on your own Mac. This page explains what “local” should mean, what it costs you, and who it is for.

What “local” should mean

“Local” gets used loosely. Some apps record locally and transcribe in the cloud. Some keep the transcript on your device but send it out to write the summary. Some store the audio locally and the notes in a web account.

A local AI note taker, in the full sense, does all four steps on your computer:

1. Capture. Your microphone and the computer’s audio are recorded on the device.
2. Transcription. Speech to text runs on the device.
3. Summaries and answers. The language model that writes the summary and answers your questions runs on the device.
4. Search. The index you search later lives on the device.

If any of those steps needs a server, the meeting leaves your computer. Ask which of the four a tool does locally before you trust it with confidential calls.

How Hearkeeper does each step on the Mac

Hearkeeper is a Mac app, not released yet, built so that all four steps run on your computer.

Capture. It records your microphone and your computer’s audio directly on the Mac. No bot joins the call and nobody else installs anything. It works with Zoom, Google Meet, Microsoft Teams, Slack huddles, Webex and FaceTime, and with meetings in the room.

Transcription. Hearkeeper uses whisper.cpp, accelerated with Metal, Apple’s graphics framework. You get live captions with about 2–4 seconds of delay during the meeting. After you stop, a final, more accurate pass with Whisper large-v3 turbo replaces them and separates the speakers. On an Apple silicon Mac we measured that final pass at about ten times faster than real time: an hour of Icelandic in under six minutes.

Summaries and answers. A local language model writes the summary, the decisions and the to-dos. No API key is needed. You can ask questions during the call with Command-Return, and across all your meetings afterwards. Every line and every answer links to the second it was said, so you can check it against the recording instead of trusting the model.

Search. Full-text and semantic search, stored in one SQLite file on your disk.

Transcription covers 99 languages. Icelandic is first-class, and meetings that mix Icelandic and English are detected automatically; see Icelandic transcription.

It works offline

Because nothing depends on a server, Hearkeeper works without a connection: on a plane, in a building with bad Wi-Fi, or on a network you do not trust. You record, and the transcript and notes are made on the spot.

Nothing is uploaded unless you choose to connect something: a team server your organisation runs, or an AI service with your own API key. Both are opt-in. If you connect neither, the meeting stays on your Mac.

Your files stay yours

Local is only half of it. The other half is that you can get your data out.

• Audio and notes are your own files on your own disk. You can back them up, move them or delete them like any other file.
• Notes export to Markdown, a plain text format that any editor can open, now and in twenty years.
• Deleting is simple. If you never connected a server, a meeting deleted on your Mac leaves no cloud copy behind, because there never was one.

The trade-offs

Running models on your own computer is not free of costs, and it is better to know them up front.

• You need a reasonably recent Mac. Hearkeeper needs macOS 14.2 or later. It runs on Apple silicon and Intel Macs. The speed figure above was measured on Apple silicon; expect an older Intel Mac to be slower.
• The models are a download. Speech and language models take real disk space. Hearkeeper shows you the size before it downloads anything.
• Your Mac does the work. Transcribing and summarising uses the processor while it runs. The app stays usable while the final pass runs in the background, but a long meeting on a slower Mac takes longer to finish.
• A local model is not the largest model. A language model that fits on a laptop is smaller than the largest models running in a data centre. That is why every line links to its source moment, and why you can connect your own API key if you decide a particular meeting may leave your Mac.
• Backups are your job. If the meetings live only on your Mac, include them in your backups.

For many people these are easy trades. For some, the cloud is the better fit, and that is fine.

Who local meeting notes are for

A local note taker makes the most sense when the content of the meeting is the sensitive part:

• Lawyers in client meetings, where confidentiality is a professional duty.
• Therapists, doctors and counsellors, whose conversations are among the most private there are. Check your professional body’s rules on recording and notes before you start.
• Journalists recording interviews, where protecting a source can mean not handing the audio to a third party.
• HR, finance and leadership teams discussing people, results or deals before they are public.
• Researchers interviewing participants under an ethics approval that limits where data may go.
• Anyone who would simply rather their conversations not sit on someone else’s server.

Local processing does not replace consent. The people in the meeting still need to know they are being recorded. Hearkeeper starts every recording with a consent step (read a notice aloud, post it in the chat, or both), saves a consent receipt with the recording, and keeps a recording pill on screen that shows up in screenshots and screen shares and cannot be hidden. The page on recording meetings with consent goes into more detail.

How to check any “private” note taker

Before you use any tool for confidential meetings, including ours, ask:

• Which of capture, transcription, summaries and search run on the device?
• Does it work with the network turned off? Try it.
• Where is the audio stored, and for how long?
• Can I export my notes in a plain format, and delete everything?
• How do the other people in the call know they are being recorded?

The website’s own data handling is described on the privacy page.

Join the waitlist

Hearkeeper is coming to the Mac first. If you want meeting notes that are made and kept on your own computer, join the waitlist and we will email you when there is a build for you.