Ampersand
0.4.0

What changed

• Ampersand can now record a meeting — in person or on a call — and turn it into a note with a summary, the decisions, the action items as real checkboxes, and the full transcript folded into a toggle underneath. Type /record in the note you are already writing and the recorder appears where your cursor is, or start one from the menu bar or ⌥⌘R when there is no window in front of you.
• The speech is transcribed on your own Mac. A Whisper model runs locally on the GPU — downloaded once, around 570 MB, or 140 MB for a faster and rougher option — and the audio is never uploaded to us or to anyone else. It is kept outside your vault so nothing syncs it, encrypted at rest with a key belonging to that machine, and deleted as soon as the transcript exists unless you choose otherwise. Settings → Meetings lists every recording still on your Mac, with its size, and lets you delete any of them.
• No bot joins your call. Nothing appears in the participant list and nobody has to approve anything — Ampersand captures the audio your Mac is playing, so it works with Zoom, Meet, Teams, FaceTime and anything else that makes noise, and it works for a conversation in a room, which no meeting bot can do. On a call your microphone and the far end are separate tracks, so the transcript can say who spoke without guessing. In a room everyone shares one microphone; rather than guess badly, the transcript is left unlabeled and the note says so.
• The one part that can leave your Mac is the transcript, and only if you have pointed the AI summary at a cloud provider — the same bring-your-own-key setup as every other AI feature. Settings names that provider in plain text whenever the one you have chosen is off-device, and says nothing when it is local. Choose Apple Intelligence or a local model through Ollama and nothing leaves the machine at any point.
• The summary streams into the note as the model writes it, so stopping a recording is followed by words appearing rather than a minute of nothing. One undo removes the whole thing if you want to start over.
• A wrong language guess used to be silent and expensive: speech recognition does not fail on a language it misreads, it translates, so a Serbian conversation came back as fluent English. Detection now runs once per recording against the longest stretch of speech it has, and Settings → Meetings has an explicit language picker for when it still gets it wrong.
• Optionally, on calls, Ampersand can sample occasional still frames of the shared screen, read the text on them on your device, and put them in the note — useful when someone walks through a dashboard. It is off by default and a separate choice from the audio, because a screenshot of somebody else’s screen often carries more than the conversation does.
• Recording a meeting is part of Pro. Reading a meeting note never is — it is an ordinary markdown file in your own folder, and every note you record keeps opening whatever happens to your license.
• Whether you may record a conversation, and whether you have to tell the other people first, depends on where you and they are. Ampersand records the way a voice recorder on the desk would: you are the one operating it, and macOS shows its own orange microphone indicator the whole time.

Ampersand updates itself in place, so you are probably already on a newer build than this one. If you are installing fresh, the download page always has the current release for macOS, Windows and Linux.