The meeting-notetaker landscape, mapped by someone who churned through six

I have installed and then uninstalled six meeting notetakers since last autumn. Not to review them for a living. Each one failed me in some specific, boring way, and I went looking for the next.

This is not a ranking, and it deliberately carries no prices. Prices change, features ship, and half of what I could tell you about any given tool would be stale by the time you read it. What does not change is the shape of the market, and the shape is the thing I needed before I could stop churning.

The first fork is architectural, and once you see it you cannot unsee it: does a bot join your call, or does something capture the audio on your own device. Everything else hangs off that one choice.

the bot camp

Otter, Fireflies, Fathom, tl;dv, and a long tail behind them. The pattern is shared: a bot dials into your Meet or Zoom or Teams call as a participant, records the stream, and does the processing in the cloud. Reputations differ at the edges: some lean toward sales and CRM, some toward personal notes. Architecturally, they are one animal.

What you get is strong integrations and calendars wired in, with someone else's servers doing the heavy lifting. What you pay is a visible robot in the participant list and your meeting audio leaving the building. For plenty of teams that trade is completely fine. For one of my Brno clients, whose legal team treats call recordings as regulated data, it was a non-starter before the first demo even loaded.

the on-device camp

Jamie is the one I keep returning to here. No bot joins the call; it captures system audio on your own machine, which means it works the same whether the call is on Teams or happening in a room with no software running at all. The company is Berlin-based and leans hard on the EU-data angle, which is exactly what the bot-averse client cared about.

Granola sits in a slightly different spot: an augmented notepad rather than a transcript machine. You type sparse notes, it captures audio locally and fills in around them afterward. Krisp arrived from noise-cancellation and added bot-free notetaking on top. Different origins, same instinct. Keep the capture on the device.

the DIY route

Then there is the route I actually started from: rolling my own with Whisper on the old 3090 in the closet. No vendor, no cloud, audio that never leaves my LAN. It is the most private option by a wide margin and the most work by a wide margin. Diarization, the who-said-what part, is mediocre. That's true everywhere, not just in my garage, and it earns its own paragraph later.

how I actually pick now

For the first three tools I chose on summary quality. I would run a call through two of them, read both write-ups, and keep whichever read better. Wrong axis. It took me three tools to notice, because the summaries are all roughly the same: the models underneath are similar and the meetings are not Shakespeare.

Here is the order I use now, most-binding constraint first:

• Can this client's audio legally and contractually go where this tool would send it. If no, the tool is out, however nice its output.
• Does it cover in-person meetings, not just video calls. A good share of my real decisions happen in a room in Ostrava, not on a screen.
• Does the output land somewhere my task tracker can reach, so the action items do not die in yet another app I never open.

The tool you pick barely writes a better summary than the next one; it wins or loses on whether it joins reliably and where the audio sleeps at night.

the boring stuff that actually decides it

Everything interesting about these tools is marketing. Everything that made me switch was dull. One joined maybe four calls out of five and I could never predict which one it would flake on. Another was 38 minutes into a 90-minute board call before anyone noticed it had quietly dropped. A third routed my audio through a region my client had explicitly ruled out, which I only caught by reading a settings page nobody is meant to read.

So that is the map. The camps are real, the architectural fork is real, and past that the differentiator is unglamorous reliability and where the bytes go.

Pick for the constraint that would get you fired, not the summary that reads nicest.