What actually leaves your Mac
Three questions to ask any meeting notes app, and Listenr's answer to each.
Three questions tell you what a meeting notes app does with your calls: where the audio becomes text, what the notes model is sent, and what a share link exposes. For Listenr the short answers are: transcription runs on your Mac unless you pick OpenAI’s cloud backend; the notes prompt holds more than the transcript but no email addresses; and a share link is public to anyone holding it, never expires, and ends only when you remove the uploaded copy.
Where does the audio become text?
Listenr records two things: your microphone, and the call audio, which a Core Audio process tap picks up from each app before it reaches your speakers or headphones. That needs the Microphone and System Audio Recording permissions. Screen Recording is optional. By default the tap hears every app except Listenr, so music or a notification sound can end up in the transcript. A setting restricts it to meeting apps and browsers.
There are five transcription backends, and four of them run on the Mac:
- Apple is the default. It uses the operating system’s on-device speech engine, downloads nothing, and transcribes one language that you set.
- Parakeet Ultra (the recommended download, about 632 MB), Parakeet v3 and Parakeet v2 run on the Neural Engine. Ultra and v3 cover 25 European languages and detect which one is spoken.
- SenseVoice covers 50+ languages, including Chinese, Japanese and Korean.
- Whisper runs locally through whisper.cpp, which you install with Homebrew.
- OpenAI is the one cloud backend. It sends roughly ten-second windows of meeting audio to OpenAI with your own API key.
Sizes and language coverage are in which local model your Mac can run.
What is the notes model told?
Every notes provider gets the same prompt:
- the meeting title, date, duration and platform
- participant display names, with join and leave times when known
- your own notes in full, plus attached images for providers that accept them
- the meeting chat, when the Zoom reader or the Meet extension captured it
- the transcript, one line per speaker turn, labelled You, Speaker 1 or a name. Over 120,000 characters, the first and last 30% are kept and the middle is dropped
Participant email addresses are not in it. Your choice of provider decides where the prompt goes, not what it contains.
| Notes provider | What leaves the Mac |
|---|---|
| Local model (MLX) | Nothing. The model is downloaded once from Hugging Face |
| Apple Intelligence | Nothing |
| Your own Anthropic key | The prompt, from your Mac to Anthropic |
| Your own OpenAI key | The prompt, from your Mac to OpenAI |
| Listenr Cloud (the default, Pro) | The prompt, through Listenr’s backend to the model Listenr hosts |
For Listenr Cloud, the backend stores a monthly token count per account and not the prompt or the notes. The prompt is still in that server’s memory while the notes are written. If that matters to you, a local model or your own key keeps Listenr’s servers out of the path, and both work on the free plan.
Where do speaker names come from?
Not from your screen. Listenr does not capture or OCR the meeting window.
Voices are separated on your Mac, and when the meeting ends they are labelled Speaker 1, Speaker 2 and so on. Your microphone is You. Names are added from these sources:
- Zoom desktop app: with “Read Zoom’s participants through Accessibility” switched on (off by default), Listenr reads the participant list, mute states, the active speaker and the open chat. It is read-only and needs the Accessibility permission.
- Google Meet: the Listenr browser extension for Chrome, Brave, Edge, Arc and Safari reads the People panel, speaking tiles and chat on meet.google.com. It does not read audio, video or captions, stores nothing, and sends only to Listenr on the same Mac. Once installed, it is used by default.
- Microsoft Teams: Listenr names speakers in Teams calls too.
- What people say: an optional pass (off by default) in which the local notes model looks for spoken names such as “thanks, Anna”.
Zoom in a browser and other apps (FaceTime, Slack, Webex, Discord) stay at Speaker N unless that last pass names someone. If you grant Screen Recording, Listenr uses it only to read window titles for meeting detection and as a fallback audio path.
What does a share link expose?
Share links are a Pro feature, also included in Team. A share link is a public URL with a random 128-bit token. Anyone with the URL can open it, with no sign-in. The page shows the title, date and time (UTC), participant names and their email addresses, the AI notes, your own notes, and the transcript. Email addresses are left out of the notes prompt but included on a share page.
The page is served with noindex, nofollow and noarchive, a no-referrer policy so the token is not passed to sites a reader clicks through to, and no-store caching. Requests are limited to 60 a minute per IP, and an unknown token gets the same 404 as a removed one.
Before you share, know that:
- Links do not expire, and there is no revoke button or command.
- Each share uploads a new snapshot, so later edits on your Mac do not reach it.
- Deleting the meeting on your Mac leaves the uploaded copy and its links working.
- Deleting the uploaded copy (
listenr team remove <id>in the CLI) ends every link to it.
Recordings are never part of a share.
What stays on your disk?
Transcripts, notes, participants, chat and note images live under ~/Library/Application Support/Listenr/, in directories only your account can open (0700, files 0600). API keys and tokens are kept in the macOS Keychain.
Meeting audio is not kept by default. A temporary recovery copy protects against a crash and is deleted once the meeting is transcribed. It is kept only if transcription degraded, so the meeting can be transcribed again. Recovery audio older than 14 days is deleted. If you turn on “Keep meeting audio”, or choose Keep for one meeting, the recording stays on your Mac.
Those permissions stop other accounts on the same Mac. They do not stop a process running as you, or root. The app is not App-Sandboxed. It runs with the hardened runtime and is notarized.
Because the library is a SQLite file plus Markdown export, a script or an assistant can read it without uploading anything.
Nobody on the call is told
No bot joins, so nobody sees a new participant or a recording banner. The Zoom reader and the Meet extension only read. Listenr has no consent prompt or announcement, and recording rules differ by country and by state, so telling people is up to you.
The three questions, for any tool
- Where does the audio become text? Which engine, does it run on the device, and what is the default?
- What is in the notes prompt? Ask specifically about names, email addresses, chat and your own notes.
- What does a share link expose, does it expire, and can it be revoked?
The same questions are scored across tools on the requirements page.