MidMeeting

Live meeting AI, explained

AI during the meeting, not after it

Most notetakers hand you a summary once the call is over. Live help means four different things in 2026, from a transcript you can read to a note you did not ask for, and here is what each vendor ships.

Mathieu-Philippe Bourgeois 10 min read

Most AI notetakers are named wrong. They do not take notes during the meeting. They record it, and when you press Stop they write a summary of a conversation that is already over. Jamie's docs say processing "typically takes about 5 - 10 minutes" once the meeting ends.1 Circleback's one in-meeting feature is a private scratchpad for your own typing.2 tl;dv lets you drop a timestamp while people talk, and that is the whole live story.3 Granola promises notes, action items and follow-ups "ready the moment the meeting ends".4 Fathom's older bot-based experience produces its transcript and summary after the call. Its new desktop app, on Mac with English-language meetings for now, is the exception, and it gets its own line below.5 Notion starts generating its summary when you select Stop, and needs at least one minute of audio to do it.6

None of that is a criticism. A good summary of a finished meeting is a real product. But it solves the archive problem. It does nothing for the sentence being said right now, the one where someone agrees to a date nobody can hit, or repeats a number that was different twenty minutes ago. That is the moment I built MidMeeting for, and it turns out "live" means at least four different things once you read the help centres.

What live means, in four tiers

01A transcript you can read

The lowest tier is text that appears while people speak. Otter's Notetaker can post a link to a live transcript into the meeting chat, a few minutes after it joins, if you switch that setting on.7 Fireflies Live Assist shows a live transcript with speaker names and timestamps. Its guide says the Fireflies bot must be in the meeting, then lists the Chrome extension and desktop app as working without the bot.8 Krisp has a Transcript Only mode you switch on from the meeting controls.9 Read's live dashboard starts transcribing once anyone speaks, after "several seconds to initialize".10 Granola has a live transcript panel next to your note, grey bubbles for system audio and green for your microphone.11 Teams has its own transcription, which is what Copilot reads.12

This tier is useful. I read the live text more than I expected to, mostly to recover a figure I half heard.

02A question box you drive

The second tier lets you type a question about the meeting so far. In Teams, each licensed participant can privately prompt Copilot during the call for who said what, where people disagree, and follow-up tasks, all in real time.12 Zoom has Meeting questions such as "Catch me up" and "What are the action items", which can only be typed from the desktop app. Other participants see the AI diamond change colour when it is active.13 Google's Ask Gemini in Meet answers from the captions plus your Workspace files, keeps the answers private to the person who asked, and stores no caption data after the meeting.14 Fireflies Live Assist ships four default suggestion buttons: Follow Up, Catch Up, Summarize So Far and Action Items.8 Granola's single-meeting recipes can run during the call; its docs give "what did I miss?" as the example.15

This is reactive help. It is only as good as the question you think to ask, and the moment you most need it is the moment you do not know you are missing something.

03A rolling summary

The third tier keeps a summary up to date without being asked. Fathom's new desktop experience includes a live summary "that updates in real time as your meeting goes on".5 Gemini in Meet has "Summary so far" for late joiners, though not on Android.16 Read produces bullet summaries "with updates roughly every minute or so".10

04Answers you did not ask for

The fourth tier is the one that interests me. The tool watches the transcript and decides on its own that something deserves a line. Cluely's Live Insights detects "real-time questions, keywords, and suggestions" from the transcript and shows them in its widget.17 Its homepage sells the product as "completely undetectable".18 Fireflies has Dynamic Topic Suggestions, but only on paid plans; free users get the four default buttons.8

MidMeeting sits in this tier.

What each tool does while the meeting is still running, from the vendors' own help pages, or a product page where the cell says so, checked 6 September 2026. Not checked here means I did not read a page about that feature.

Tool Live transcript You ask, it answers Rolling summary Notes you did not ask for Seen by the others
Otter.ai Yes, link in chat7 Yes, Otter Chat Q&A Listed, not described; summary posted at the end197 Not documented Bot in the list for these chat features; the desktop app can also record without the bot20
Fireflies.ai Yes, Live Assist8 Yes Live AI Notes update as people talk; Summarize So Far button Dynamic Topic Suggestions, paid plans Bot in the call, or none with the desktop app or Chrome extension
Fathom Vendor says yes for the new desktop app; the page only describes a live summary5 No Yes, new desktop app5 No Bot in the older mode and in the new app's Audio & Video mode; its Transcript Only and Audio Only modes are bot-free21
tl;dv No, timestamps only3 No No No Bot for scheduled meetings22; the desktop app can record without one23
Granola Yes, panel11 Yes, recipes during the call15 No, notes at the end4 No Nothing joins24
Krisp Yes, Transcript Only mode9 No Not documented No Nothing joins by default
Zoom AI Not checked here Yes, desktop app only13 Catch me up on request No AI diamond changes colour
Gemini in Meet Captions, which Ask Gemini reads14 Yes, private to you14 Summary so far16 No Platform feature, no bot
Teams Copilot Yes, Teams transcription12 Yes, private prompts Not checked here Offers a catch-up if you join over 5 minutes late25 Platform feature, no bot
Notion AI Meeting Notes Product page says real-time; the help centre does not describe it26 No No, after Stop No Nothing joins
Cluely Yes18 Yes Not documented Yes, Live Insights17 Hidden by design; screen-share hiding is $149.99 a month billed monthly27
Read.ai Yes, dashboard10 Yes, desktop app28 Yes, about every minute No Bot in web mode, none in desktop mode29
Jamie No, scratchpad only30 No No No Nothing joins31
Circleback No, scratchpad only2 No No No Bot first when joining preferences are on; the desktop app records without one32
MidMeetingthis site Yes, on your computer Yes, highlight a span No, summary after Yes, advisor notes Nothing joins; a window you can show

Reactive or proactive, hidden or shown

Two axes hide inside that table. The first is who starts the exchange. Reactive tools wait for you to ask: Copilot, apart from a catch-up offer when you join late, Ask Gemini, Zoom's question box, Fireflies' suggestion buttons, Granola's recipes. Proactive tools watch the words and decide for themselves: Cluely's Live Insights, Fireflies' topic suggestions, MidMeeting's advisors. The whole craft of a proactive tool is in how often it stays silent.

The second axis is whether the other people on the call can see the help. At the loud end, Otter's Notetaker sits in the participant list and, if you toggle it on, sends the questions and answers to everyone.

Send the questions and AI-generated answers from Otter Chat directly to your meeting for all participants to view.

Otter help centre, Manage your Notetaker settings, checked 6 September 20267

At the stealth end is Cluely, which never joins the call, shows nothing to the others, and charges $149.99 a month, billed monthly, for the plan that is "Completely hidden to meeting screen sharing software".27

Cluely takes perfect meeting notes and gives real-time answers, all while completely undetectable.

cluely.com home page, checked 6 September 202618

The platform tools sit in between. Teams and Meet do not add a participant, and what Copilot or Gemini tells you stays on your screen.1214 They are private, but they are not hidden: the platform knows, and in Zoom's case the diamond changes colour for everyone.13

Three to five seconds, not 300 milliseconds

Latency is where live tools are most tempted to round down. Here is what happens between a sentence ending and a note landing, against a bot notetaker.

When the help arrives: a bot notetaker against MidMeeting Two timelines. Top, a bot notetaker: the sentence is said at zero seconds, audio streams to the vendor for the whole meeting, and a summary arrives only after Stop. Bottom, MidMeeting: the sentence ends at zero seconds, its text is on screen a few seconds later, an advisor reads it and usually passes, and when it has something to say a note lands three to five seconds after the sentence ended. A bot notetaker audio goes to the vendor; the summary comes after Stop the meeting keeps going 0 s: the sentence is said audio streams to the vendor the whole time Stop summary arrives minutes after the meeting MidMeeting on your own computer; the note lands while the sentence is still warm 0 s 1 s 2 s 3 s 4 s 5 s sentence ends text on screen whisper, on this machine advisor reads it PASS (most sentences) a note lands on your screen forty words or fewer, with the reason 3 to 5 s after the sentence ended A bot notetaker summary after Stop 0 s: the sentence is said the meeting keeps going audio streams to the vendor the whole time Stop summary arrives minutes after the meeting MidMeeting 3 to 5 s after the sentence 0 s 1 s 2 s 3 s 4 s 5 s sentence ends whisper, on this machine text on screen advisor reads it: PASS (most sentences) a note lands on your screen only forty words or fewer with the reason
The two lanes share a starting point and nothing else: the top one is measured in minutes, the bottom one in seconds.

A sentence ends. whisper.cpp, compiled into the app and running on your GPU or CPU, turns the audio into text and puts it on screen, a few seconds behind the speaker. That text goes to the model you connected, and when an advisor has something to say a card lands three to five seconds after the sentence ended.

I am not going to promise less than that. Cluely's homepage claims a 300 ms response time for its live transcription.18 A round trip to a language model, on top of local transcription, on top of waiting for the sentence to actually finish, does not fit in 300 ms on the hardware people own. Three to five seconds assumes the live text is keeping up. On a laptop with no GPU, the live text lags further and the cards with it. What I checked, and what I will stand behind, is that a note arriving five seconds later is still inside the conversation. The person is still on that topic and you can still ask. A summary that arrives after Stop is not inside anything.

Who said what is a guess

Every tool that names speakers from a single mixed audio stream is guessing. The guess is called diarisation, and the vendors are candid about its limits once you get past the product page. Read's desktop mode captures audio locally, and Read says plainly what that costs.

Read AI does not know who is speaking. It produces a transcript without speaker attribution.

Read help centre, How consent works on web mode vs desktop mode, checked 6 September 202629

Notion says its speaker labels work "best for virtual one-on-one meetings", where it can separate your microphone from your system audio.6 Jamie diarises automatically, then asks you to name each speaker after the meeting by listening to an audio clip, and warns that the speaker count can be wrong with poor audio.33 Granola, by default, shows Me and Them, split by microphone and system audio.11 Read's help centre shows what the guess looks like when several people share one conference-room microphone. They come out as "Conference Room - Speaker 1" and "Conference Room - Speaker 2", and the report owner types the names in afterwards.34

MidMeeting takes the Granola position and stops there. It records two tracks. What the microphone hears is labelled You. What comes out of the speakers is labelled Them. The label is which device heard the word, and it never carries a name. Them is everyone else. Several people sharing your room all land on You, so a track is a side of the call, not a person. I took that trade on purpose: a live note that says "you just agreed to Friday" has to be right about the "you", and a track label is not a guess. A name from diarisation is a guess that is usually right, which is a different thing.

What an advisor actually does

An advisor is a prompt. The four that ship, Engineer, Devil's advocate, Researcher and Note Taker, are prompts I wrote. A custom one is a prompt you write, in plain words. "You are our finance lead. Flag any number that changed since it was last said" is a complete advisor.

Every time someone finishes a sentence, the app checks which armed advisors are due. An advisor is consulted at most once per cadence window, and the Chattiness setting picks the window: Relaxed every 90 seconds, Balanced every 45, Chatty every 20. The four defaults ship on Balanced. A consultation sends the advisor its role prompt, the last ten minutes of transcript and a running summary the app keeps for the advisors and refreshes every two minutes. You do not see that summary; the one you read is written after the call. The reply is either PASS or a card, and PASS is the default. A card has a kind (claim check, risk, question, idea, correction), a text of forty words or fewer and a one-line reason. If you want a note now, a Jump in button asks the advisor for one.

[14:02:03] Them: Friday is fine for the migration, we can absorb a vendor slip.
[14:02:09] You: Good, let's lock it.
[14:02:13] Devil's advocate · risk: Nobody has said who owns the rollback if the vendor slips. Ask now, not in the retro.
           Reason: "absorb a vendor slip" names a risk without an owner.

The reactive tier is there too: highlight a span of the transcript, type a question, and the answer lands in a thread on that line. After the call, the Summary tab writes the summary, the action items or a full report you can save as HTML.

All of it runs from your own computer. Transcription never leaves the machine. The advisors and the questions send transcript text, and only text, to the provider you chose. That can be an API key from OpenRouter, Anthropic, OpenAI, Google Gemini or xAI. It can be an OpenAI-compatible endpoint such as LM Studio or Ollama on your own machine. Or it can be a coding agent already installed, through the agent bridge. Set the copilot to Off and nothing is sent anywhere; you have a plain live transcriber.

The trade-offs

  • An advisor that stays quiet feels like nothing is happening. PASS is the default, and in a well-run meeting you can go a long stretch without a card. People used to a chatty assistant read that as broken. It is not, but I have had to explain it more than once.
  • It needs a model. Recording and the live text need no account. The advisors, the questions and the summary need an API key or a local model, and a local model needs a machine that can run one. MidMeeting sells no AI and proxies nothing, so this is your setup to do.
  • It needs your computer in the call. There is no bot, so it cannot record a meeting you skipped, and there is no phone app. A bot notetaker can attend without you.
  • No names on the other people. Them is one track, so a three-person call reads as You and Them, and you sort out who said what yourself. A diarising notetaker will at least guess.
  • Local capture does not settle consent. The recording stays on your disk, but the people on the call still have the same rights they would with a bot in the room. Tell them. The window is designed to be shown.

Sources

  1. Summary, Jamie docschecked 6 September 2026
  2. Add your own notes, Circleback help centrechecked 6 September 2026
  3. Desktop app, tl;dvchecked 6 September 2026
  4. Granola home pagechecked 6 September 2026
  5. Does Fathom provide real-time transcription?, Fathom help centrechecked 6 September 2026
  6. AI Meeting Notes, Notion help centrechecked 6 September 2026
  7. Manage your Notetaker settings, Otter help centrechecked 6 September 2026
  8. Learn about Fireflies Live Assist, Fireflies guidechecked 6 September 2026
  9. AI Meeting Assistant, Krispchecked 6 September 2026
  10. Using Read's live meeting dashboard, Read help centrechecked 6 September 2026
  11. Transcription, Granola help centrechecked 6 September 2026
  12. Microsoft Copilot in Teams meetings and events, Microsoft Learnchecked 6 September 2026
  13. Asking in-meeting questions with the AI Companion, Zoom supportchecked 6 September 2026
  14. Ask Gemini in Google Meet, Google Meet helpchecked 6 September 2026
  15. Writing effective recipes, Granola help centrechecked 6 September 2026
  16. Take notes for me in Google Meet, Google Meet helpchecked 6 September 2026
  17. Live Insights, Cluely docschecked 6 September 2026
  18. Cluely home pagechecked 6 September 2026
  19. Otter Notetaker Overview, Otter help centrechecked 6 September 2026
  20. Otter Desktop App (Mac, Windows), Otter help centrechecked 6 September 2026
  21. How to Upgrade to the New Fathom Experience (Version 3.0+), Fathom help centrechecked 6 September 2026
  22. Why do we use a bot?, tl;dv help centrechecked 6 September 2026
  23. Bot-free recording, tl;dv help centrechecked 6 September 2026
  24. Security, privacy and data FAQs, Granola help centrechecked 6 September 2026
  25. Catch up on meetings with Microsoft Copilot in Teams, Microsoft Supportchecked 6 September 2026
  26. AI Meeting Notes product page, Notionchecked 6 September 2026
  27. Pricing, Cluelychecked 6 September 2026
  28. Read AI for desktop, Readchecked 6 September 2026
  29. How consent works on web mode vs desktop mode, Read help centrechecked 6 September 2026
  30. Recorder, Jamie docschecked 6 September 2026
  31. Quickstart, Jamie docschecked 6 September 2026
  32. Record meetings with the desktop app, Circleback help centrechecked 6 September 2026
  33. Identify speakers, Jamie docschecked 6 September 2026
  34. Can I edit text and speaker names in my meeting report, Read help centrechecked 6 September 2026

Mathieu-Philippe Bourgeois

Builds MidMeeting. Software engineer in Sherbrooke, Quebec. matpb.com

Record the next one yourself.

Three free meetings of any length, then $39 once. The recording never leaves your computer.

Download free See how it works

Keep reading

All posts
6 September 2026 · 10 min read

Why did an AI bot join my meeting?

A notetaker bot gets in because someone gave it their calendar. Here is the plumbing, the three places a recorder can sit, what Teams, Meet and Zoom did about bots in 2026, and why a visible tile is not consent.

 6 September 2026 · 10 min read

What AI meeting notes really cost

The pricing page shows the annual number; the monthly one is often double, and the free tier has a catch. I lined up thirteen vendors, then priced the alternative: buy once and pay for tokens yourself.

 6 September 2026 · 10 min read

Where does your meeting audio actually go?

Bot-free is not the same as local. Most desktop notetakers record on your computer and then ship the audio somewhere else. Three questions to put to any vendor, with the answers I found on their own pages.