Meeting notes¶

Meeting notes records a meeting through your computer's audio system, whether it is a call or a meeting held in the room with you (below), transcribes it on your computer, and saves the transcript as a Markdown file in a folder you choose. If you have an Assistant, it can summarise each transcript and file the meeting in its memory (below).

It is part of Agentry Desktop, it is free, and nothing about it is hosted: the audio never leaves your computer, and a transcript leaves it only if you give your transcripts to your Assistant, whose model provider then reads them. It is off until you turn it on.

It needs Agentry Desktop 0.6.168 or later on Linux, 0.6.172 or later on Windows and 0.6.173 or later on macOS, where it also needs macOS 14.4 or later: on an older macOS the app does not show it. A copy of the app without its transcriber, such as one you built yourself, does not show it either. Meetings in the room need 0.6.177 or later on Linux and 0.6.178 or later on Windows and macOS, and telling their speakers apart needs 0.6.184 or later. Telling apart the speakers on a call, and the settings in Settings → Meeting notes, need Agentry Desktop 0.7.8 or later.

Turning it on¶

Meeting notes has its own entry in the sidebar, below your agents. Its page lists what turning it on means before you do:

Recording others may need their consent; tell them you are recording.
While it records a meeting, it captures your microphone even when you muted yourself in the meeting app. Use Pause for anything you do not want kept.
When no meeting app can be recorded on its own, it records everything your computer plays, other sounds included, and the transcript says so. On Windows it always records a whole output device (below).
On macOS, the first time it records, macOS asks whether Agentry Desktop may use the microphone and record system audio.
It downloads a speech model the first time, about 200 MB, or about 570 MB on a Mac with Apple silicon.

Turn on Meeting notes… asks you once more in a dialog. Once it is on, the app downloads the speech model and measures how fast your computer transcribes (Transcription), and adds a Meeting notes card to your Home page. The page also offers to start Agentry when you sign in, because calls are noticed only while Agentry runs.

While Meeting notes is on, closing the window keeps Agentry running in the background, with its icon in the system tray, so calls are still noticed.

Its settings are on the Meeting notes page, above the list of meetings, and in Settings → Meeting notes. Both show the same settings, and a change made in one shows in the other. While Meeting notes is off, Settings → Meeting notes says so and has Open Meeting notes: you turn it on from the page, which tells you first what it records. Settings lists Meeting notes only where this copy of Agentry can record meetings.

Turn off Meeting notes, the last of the settings in either place, stops a recording in progress (keeping it, to transcribe once it is on again) and stops noticing calls.

Recording a meeting¶

There are three ways a recording starts:

By hand. Press Record… on the page or the Home card once your meeting has started, and agree in the dialog.
On an offer. When a meeting app starts a call, Agentry offers to record it: "Record this meeting? (Zoom)", in a notification and at the top of the page. On Linux the notification has Record and Not now buttons. On Windows and macOS, clicking it opens the offer in the app. An offer you do not answer lapses after two minutes and records nothing, and a call you answered is not offered again.
Automatically, for the apps you chose. Under Meeting apps in the settings, each app whose call Agentry has noticed can be set to Record automatically; choosing that asks you in a dialog. A browser's calls come from any site, so a browser set to record automatically records every call in it. Your Assistant can also set up recordings from your calendar, if you let it (below).

A recording is always shown, however it started: a notification that stays on screen, with Pause and Stop; a red dot on the tray icon, with the recording in its menu; and the page and the Home card. If the notification cannot be shown, an automatic recording does not go ahead.

The page has a line you can paste into the meeting's chat to tell the others: "I'm recording this meeting to keep notes. They are transcribed on my computer."

A recording stops when:

you press Stop;
the meeting app stops using the microphone for 30 seconds;
a recording window set up by your Assistant ends;
it has run for four hours, or the length you chose under Stop a recording after (1 to 8 hours);
you quit Agentry. What was recorded is kept and transcribed the next time Agentry starts.

Pause stops recording until you resume; the transcript marks the gap.

What it records¶

A recording has two tracks, transcribed separately and merged by time:

Me: your microphone.
Them: the meeting app's sound, everyone else on the call. It is not split by person unless you turn on Tell speakers apart on calls (below).

Agentry finds the call through the audio system, never through the meeting app: a call is an app that uses a microphone and plays sound at the same time. Agentry does not open your microphone before the meeting app has, unless you press Record yourself.

	Linux	Windows	macOS 14.4 or later
Them	the meeting app's own sound, each of its playback streams	everything played on the speakers or headset the meeting app uses	the meeting app's own sound
When the app cannot be recorded on its own	everything the app's output device plays	everything your default output plays	everything your computer plays, except Agentry
Me	the microphone the meeting app uses	the microphone the meeting app uses	your default microphone

Recording everything a device plays includes other sounds on it, such as another app's notifications or a video. When that happens, the transcript's header says far_side: whole-device and the recording's notification says so. On Windows, far_side: app-device means everything played on the device the meeting app uses.

Muted audio. Muting yourself in the meeting app does not mute your microphone for the operating system, so the Me track still records you. Use Pause, or mute the microphone itself.

Speakers. On speakers, your microphone also hears the other side. Agentry drops a Me line that repeats a Them line from the same moment, but headphones give a cleaner transcript.

Meetings in the room¶

For a meeting held around you rather than on a call, press Record in the room… on the page or the Home card, or choose Record a meeting in the room in the tray's menu. Agentry records your microphone alone, and nothing your computer plays, until you press Stop or the recording reaches its longest length.

Only by hand. A meeting in the room is never offered, never recorded automatically and never set up from your calendar. Agentry asks in its own dialog each time. Tell everyone in the room that you are recording.
The microphone. It records your default microphone, or the one you chose under Microphone in the settings. That choice also applies to Record when no meeting app is found; a call whose meeting app Agentry follows keeps recording the microphone that app uses. If the chosen microphone is not connected, the default is recorded, and the notification, the page and the transcript say so.
It stops when you press Stop, when it has run its longest length (4 hours, or what you chose under Stop a recording after), or when you quit Agentry, which keeps what was recorded. No meeting app ends it. Pause works as in any recording.
Still recording? When nobody has spoken for 15 minutes, a notification asks "Still recording?", with Stop and Keep going. If you do not answer, it keeps recording.
On macOS, the microphone needs the permission described below.

Speaker 1, Speaker 2 and People in the room¶

With Tell speakers apart in meetings in the room on, as it is unless you turn it off in the settings, Agentry tells the voices apart after the meeting is transcribed, on your computer, with sherpa-onnx's speaker diarization, which Agentry Desktop carries. The first time, it downloads about 46 MB of models (pyannote segmentation 3.0 and NeMo TitaNet small) from Hugging Face at a fixed revision, checked against the SHA-256 Agentry pins, like the speech models.

Record in the room… on the page asks People in the room (optional), a number from 2 to 20:

With a number, the voices are split into at most that many. Someone who said little may not appear.
Left empty, the number is estimated, so one person may appear under two numbers. A recording started from the Home card or the tray does not ask, and estimates.

The labels mean:

Speaker 1, Speaker 2, … are voices, numbered by who spoke first, the same number for the same voice within one transcript. They are not people. Two people with similar voices may share a number, one person may appear under two, and the numbers in one meeting have nothing to do with another meeting's. Agentry never puts a name on them. You are one of the voices, and nothing says which.
Room is everyone in the room, not split by person. Every line says Room when Tell speakers apart is off, when this copy of Agentry does not carry the speaker separation, or when it fails; the transcript says which, and why.

Crosstalk, a quiet speaker and a large group are where it most often puts a word to the wrong voice. The meetings list shows a meeting in the room as "In the room", with how many speakers were told apart.

When your Assistant summarises a meeting in the room, it keeps these labels. It writes a name only where the transcript gives one, such as someone saying their own name, and marks it as inferred: "Speaker 2 (Dana, from the transcript)" (Meeting notes on its page).

Telling speakers apart on calls¶

Tell speakers apart on calls, in the settings, is off unless you turn it on. With it on, a call's transcript keeps Me for you and splits Them into Speaker 1, Speaker 2, … for the others on the call. Agentry does this on your computer once the call is transcribed, with the same speaker separation and the same models as meetings in the room (above): about 46 MB, downloaded the first time either one needs them. It adds a few minutes to each transcription.

Me stays Me. Your microphone is its own track, so its lines are never split.
The Speaker numbers mean what they mean in the room: voices, numbered by who spoke first, not people. Two people with similar voices may share a number, and Agentry never names them. A call does not ask how many people took part, so the number is always estimated, and one person may appear under two numbers.
A far side that is a whole device (far_side: whole-device, or app-device on Windows) includes the other sounds that device played, so the Speaker lines may carry them too. The transcript says so.
When it cannot run, because its models could not be downloaded or the separation failed, the far side stays Them, and the transcript says why, for example: "Telling the far side's speakers apart failed (the speaker separation stopped with an error), so every far-side line is Them."

The meetings list shows a call told apart this way with how many voices were heard on the far side, such as "Zoom, 3 other speakers".

What you are asked on each system¶

Every system. Agentry asks you in its own dialog before turning Meeting notes on, before each Record you press on the page, and before an app joins your automatic list.
Linux. Nothing more: the audio system does not ask.
Windows. Windows lists Agentry Desktop among the apps using your microphone while it records, as its microphone icon shows. Windows' privacy setting for desktop apps' microphone access applies.
macOS. macOS asks, the first time, whether Agentry Desktop may use the microphone and whether it may record system audio. If you declined, macOS gives Agentry silence instead of an error. So when a side of a recording has been completely silent for 30 seconds while it should hear something, Agentry says so, in a notification and on the page, with an Open System Settings button that opens the Microphone or the Screen & System Audio Recording pane.

Transcription¶

When a recording stops, it is transcribed in the background on your computer by Whisper, through whisper.cpp, which Agentry Desktop carries with it. A track that is silent is not transcribed. Voice detection skips the silence in the rest.

Language. Under Language in the settings:

English is more accurate for English.
Multilingual detects the language spoken.

Models. Agentry downloads a model the first time it needs one, from the whisper.cpp project on Hugging Face at a fixed revision. It checks every model's SHA-256 against the one Agentry pins, at download and before each use, and deletes a file that does not match. Models are kept in Agentry's own folder.

Your computer	Model	Download
Linux, Windows, a Mac with Intel	Small, or Base where Small does not keep up	about 190 MB, or 60 MB
A Mac with Apple silicon	Large v3 Turbo, on the graphics processor	about 574 MB
Linux or Windows, on a graphics card you chose	Large v3 Turbo	about 574 MB

Automatic picks for you. On a processor, it measures how fast your computer transcribes a minute of speech with Small, and uses Base instead if Small is slower than the meeting itself. Model in the settings lets you choose another.

Your graphics card (Linux and Windows, Agentry Desktop 0.6.175 or later). Once the model is ready, Agentry measures each graphics card it can use, by transcribing the same minute of speech on it. A graphics card is often slower than the processor at this, so a card is offered only when it is fast enough to run Large v3 Turbo, the more accurate model, as fast as the meeting is spoken. The offer reads "Use your graphics card with the better model (574 MB)", with Use the graphics card and Keep the processor. Either answer is kept, and the offer is not made again.

Transcribe on, in the settings, chooses the processor or the fastest card Agentry measured, and says whether that card keeps up; you can choose a card that does not, and transcripts then take longer than the meeting. Measure the graphics card again is for a new card or driver. If a transcription on the card fails, it is done again on the processor.

The transcript¶

Each transcript is a Markdown file named for the meeting's date, time and app, such as 2026-10-26-1400-zoom.md (the event's title instead of the app for a recording your Assistant set up), written to Documents/Agentry/Meeting notes/ unless you chose another folder under Transcripts folder. Only the folder picker can change the folder. A transcript is written whole and never replaces another file.

The file starts with a YAML header: when the meeting started and ended, the app, the language, the model, what Them holds (far_side) and who took part (participants, unknown unless the calendar said). Then comes a note on what the labels mean, and one line per turn:

[00:00:12] Me: Shall we start with the release dates?
[00:00:15] Them: Yes. The beta is ready on Thursday.

A meeting in the room is named …-meeting-in-the-room.md, has no Me and no Them, and its header adds setting: in-room, far_side: none and the microphone recorded. When the speakers were told apart it says diarization: done, whether the number of people was given or estimated (speaker_count), which models told them apart (diarizer), and lists Speaker 1, 2, … under speakers; otherwise diarization is off, unavailable or failed, and every line is Room:

[00:00:01] Speaker 1: Shall we start?
[00:00:04] Speaker 2: Yes, let's.

A call keeps Me and Them unless Tell speakers apart on calls was on, and then its header has diarization too. When the far side was told apart, it says diarization: done, speaker_count: estimated and the diarizer, and lists Me and Speaker 1, 2, … under speakers:

far_side: app
diarization: done
speaker_count: estimated
diarizer: sherpa-onnx v2.0.1, segmentation pyannote 3.0, voice embedding NeMo TitaNet small

[00:00:01] Speaker 1: Shall we start?
[00:00:04] Speaker 2: Yes, let's.
[00:00:08] Me: I agree.
[00:00:12] Speaker 1: Good.

When it could not be, diarization is unavailable or failed, and every far-side line is Them.

The time on each line counts from the start of the recording. The header's gaps lists the pauses, and silent a side that recorded nothing. A recording your Assistant set up takes the calendar event's title and invitees; participants is then who was invited, not proof of who spoke. Words, names and numbers may be wrong: the transcript is made by a machine.

The meetings list on the page, below the settings, shows each meeting with Open transcript, Show in folder, Transcribe for one that failed, Delete recording and Remove from list. Removing a meeting from the list leaves its transcript in the folder.

Recordings and how long they are kept¶

A recording's audio is deleted once its transcript is written. With Keep recordings for 7 days on, the audio stays in Agentry's own folder for seven days, never in the transcripts folder, and you can delete it sooner. A recording whose transcription failed or is waiting keeps its audio for up to seven days, so it can be transcribed again. Agentry removes expired audio when it starts and every hour.

Giving transcripts to your Assistant¶

If you have an Assistant, the settings and its Configuration's Meeting notes card have Give my meeting transcripts to your Assistant, off until you turn it on in a dialog. With it on:

each transcript is also copied into the Assistant's share folder, under meetings/;
an Assistant that is running is asked to summarise it, with its meeting-notes skill;
an Assistant that is busy is asked once it is free;
an Assistant that is stopped is never started for this: the transcript waits in its share folder, and the meeting's row offers Ask your Assistant to summarise.

Ask me first leaves every transcript in the share folder until you ask. What the Assistant does with a transcript is on its page.

Recordings your Assistant sets up¶

Your Assistant can read your calendar and set up a recording for a meeting you asked it to record, if you switch on its Can set up meeting recordings entry (Meeting notes). A recording it sets up is a window around the meeting, from 5 minutes before it starts to 15 minutes after it ends. The window records only a call Agentry notices while it is open, from the app the meeting is in or from a browser; a meeting you skip records nothing. Each window is listed on the page and the Home card with Cancel, and a recording made in one is shown as any other, with the calendar's title.