Local-first meeting notes · Windows
MIN records both sides of a call without a bot joining, transcribes it on your own computer as the meeting happens, and turns the notes you typed into a clean write-up. No account. No API key. No subscription to this. After the one-time model download, recording and transcription never touch the network. MIN itself uploads nothing and has no account anywhere; point it at a folder your cloud already syncs and your meetings reach your other computers, because that program does the uploading.
~117 MB installer · unsigned, so Windows will warn you · free and open source · MIT
How it works
A summary written by something that was not in the room reads generic. MIN starts from your notes and uses the transcript to fill them out.
That claim is not yet proven. The test is a blind reader trying to tell a write-up built from your real notes apart from one built from a different meeting's notes. It has never been run.
01 · Record
Your microphone and the call audio are captured as two separate tracks. Nobody sees a bot join the meeting, because none does. Click a meeting on your home screen and the note opens with its title, time and attendees already in it.
02 · Type
Half sentences, a heading, a name with a question mark. Your headings become the sections of the write-up, so the structure stays yours.
03 · Write up
Transcription runs on your machine. There is no local model for the write-up: the prompt goes to the assistant you already pay for, by clipboard or into Claude Desktop through the MCP extension.
Why it is built this way
"Who said this" is just "which file was it in". A tool that works from one mixed recording has to infer the speaker; MIN keeps two recordings, so there is nothing to infer. No speaker-identification model, no clustering, no guessing, and the two tracks can never drift apart.
Transcription runs on your CPU, costs nothing, and works with the network cable pulled. It has been benchmarked on exactly one desktop, faster than most laptops, so the specs below are a ceiling rather than a promise.
Every meeting is a folder of plain markdown you can open in Notepad. No database, no proprietary format, nothing to export.
Three dots on any note, then Move to trash. It goes to the recycle bin your computer already has, so it is yours to put back. There is no "are you sure" step, on purpose: the recycle bin is the confirmation, and a dialog in front of something this reversible only teaches you to click through the ones that matter.
Audio is removed once transcription succeeds, not merely kept locally. If a transcription fails the audio is kept, because then it is the only way to try again.
The transcript fills in beside your notes as people speak: the call on the left in grey, your microphone on the right in green, already sorted by who said it. A line lands when its speaker pauses, so a short remark appears a second or two later and a long unbroken stretch takes longer. Those lines are the transcript; nothing is run over the audio a second time.
Paste the secret iCal address from your calendar settings and your next meetings appear on the home screen. No Google sign-in, no OAuth, no consent screen: it is one HTTPS address MIN reads on a timer, and it works with Outlook, iCloud and Fastmail too. That address is a credential, so it stays on your machine and is never printed in a log.
Just before a meeting on your calendar, a small card appears in the corner naming it, with a Take notes button that opens a note for that meeting and starts recording, and an X to send it away. Either answer sticks, and it will not raise that meeting again. It cannot take focus, so it never steals the call it is announcing. All-day entries never nudge you, a meeting that started five minutes ago still does, and it does not skip a meeting for having too few people in it, so a lecture or a one-to-one counts like anything else.
When Zoom, Teams, Chrome or anything else opens your microphone, a small card in the corner says Meeting detected, names the app, and offers Take notes. MIN never listens for this: Windows keeps a record of which apps are using the microphone, the one behind the microphone icon in your tray, and MIN reads that. Silence it for one app from the card and it stays quiet for that app until you allow it again in Settings. It never fires while MIN is already recording. To be there when a call starts, MIN starts with Windows and lives in the tray; closing the window keeps it there, and Quit is on the tray icon.
The audio is written into the meeting's folder while the meeting is still going, a few seconds at a time, instead of being held in memory until you stop. If the machine dies, Windows forces a restart, or MIN is killed, the next launch finds what was written and finishes it into a normal note. MIN also asks Windows to stay awake while recording, so a closed lid does not quietly stop the capture.
The transcript and your notes are written into the meeting's folder as the meeting runs, a couple of seconds behind the speech. Install the Claude Desktop extension and Claude can read the meeting so far, on the subscription you already have, with no API and no bill. Ask it what was said about the budget ten minutes in, and again at the end.
A round waveform button in the note is grey and still when nothing is captured and green and moving when it is. Switch to the call and a small pill floats above every window with the levels, the clock and a stop button. Drag it anywhere and it stays there. It cannot take focus, so it never interrupts the meeting.
Stop ends the capture, not the note. Press it again and the recording carries on into the same note as another part of it, whether that is thirty seconds or a day later. One transcript, one clock running through it, and two honest numbers: the audio that was captured, and the wall-clock window it happened in.
Search inside it and matching lines stay with the match highlighted. Click any line to copy it with its timestamp, labelled the way the bubble beside it is labelled, so your own lines read Me. The saved transcript, and the button that copies the whole transcript at once, say You instead. And a reminder to get consent when transcribing others, where the recording is visible rather than buried in a settings page.
The home screen is what is coming up and every note you have taken, newest first. One box searches titles, the notes you typed and the full transcripts, on this machine, over the same markdown files you own.
Specifications
Every number below was measured on one machine, the developer's desktop: an AMD Ryzen 5 5600G, six physical cores, 15.3 GB of RAM, CPU backend. The project's own hardware check refuses to call it a valid test bed. It puts the box at roughly 1.5x faster than the median 4-core laptop, which has never been benchmarked, and warns about that in every result file, three of the four also carrying "validTestBed": false. Read every figure as an upper bound. Raw JSON in results/.
| What | Measured on that one desktop |
|---|---|
| Transcription speed | 9.8x real time at the 30-second chunk the app uses, and about 10x to 13x end to end, so 4.5 to 6 minutes per hour of audio. Longer chunks are slower, about 5.6x at 360 seconds, and 398 seconds is a hard ceiling the recogniser aborts on. A full hour has not been run end to end, and a 4-core laptop should expect nearer 7 to 9 minutes |
| Storage, 1 000 hours | ~95 MB of markdown, calculated from transcript sizes rather than timed. Audio is deleted after transcription |
| Speaker attribution | Exact, from separate tracks. No diarization model involved |
| Installer | About 117 MB. The speech models are not in it: 642 MiB downloads once, the first time you press Record |
| Where things live | Meetings are markdown in a folder you own. Models, the search index, the MCP index, your settings and the cached calendar sit in %APPDATA%\MIN, plaintext and searchable. Uninstalling deliberately removes neither, so notes are never destroyed on the way out |
| Live transcript | Written as the meeting runs, from the two tracks, with no diarization model. Lines appear when a speaker pauses, so a short remark lands a second or two behind and a single utterance is capped at 20 seconds. It can be switched off, and then the transcript is written after you stop |
| Calendar | Optional. One secret iCal address, pasted in Settings, fetched over HTTPS every 5 to 60 minutes. No sign-in, no OAuth, no account. The next 14 days appear on the home screen, five meetings at a time, and a note opened from one of them takes its title and attendees |
| Network use | Three things, all your choice: fetching the calendar address you pasted, sending a write-up out, and, if you point the meetings folder at one your cloud syncs, that program uploading it. MIN holds no account and uploads nothing itself |
| Requirements | Windows 10 or 11, 64-bit. Transcription uses the CPU backend, so no graphics card is needed |
| Licence | MIT. Speech models are CC-BY-4.0 and MIT, fetched at runtime |
Download
MIN-Setup.exe and double-click it.Beta build, not an official release. Direct download. Every version is listed on the Releases page.