What is the best meeting notetaker?
Nine of them put against the same seven questions, with a link to the vendor page behind every answer.
The scorecard
● yes◐ partly○ no
| Product | On-device transcription | Choice of model | No bot joins | Native Mac app | Local CLI or MCP | No training on your content | Records without you |
|---|---|---|---|---|---|---|---|
| Listenr | Yes: On-device transcription. The default is Apple’s on-device engine, in one language you set. Parakeet, a download, adds 25 European languages with detection. Four of five transcription backends run on the Mac; the fifth, OpenAI, is a cloud option you have to pick. | Yes: Choice of model. A model on your Mac, Apple Intelligence, your own Anthropic or OpenAI key, or Listenr Cloud on Pro. | Yes: No bot joins | Yes: Native Mac app. Written in Swift and SwiftUI for macOS and nothing else. There is no Windows, Linux or browser version to keep in step with. | Yes: Local CLI or MCP. The listenr CLI and a local MCP server, free on every plan, reading the database on your own disk. Every command the app has is a CLI verb, and the MCP server has seven read-only tools by default and over 150 when control and settings are enabled by flag. | Yes: No training on your content. With a local model or Apple Intelligence nothing is sent anywhere. Listenr Cloud, the hosted option on Pro, stores neither the transcript nor the notes and keeps only a monthly token count per account. | No: Records without you. It records the audio your Mac is playing, so you have to be in the meeting. |
| Fireflies.ai | No: On-device transcription. Third-party transcription vendors, under a stated 0-day retention contract. | No: Choice of model | Partly: No bot joins. The desktop app records system audio with no bot; the bot remains for meetings you do not attend. | Partly: Native Mac app. A Mac build of an app that also ships for Windows, the web and mobile. | Partly: Local CLI or MCP. A GraphQL API available to every user including the free tier, capped at 50 requests a day on Free, plus MCP connectors. Hosted, not local. | Yes: No training on your content. Fireflies states your data is never used to train AI models, internally or externally. | Yes: Records without you |
| Notion AI meeting notes | No: On-device transcription. Notion names its transcription subprocessors publicly: OpenAI, Anthropic, Fireworks, Baseten, xAI and AssemblyAI. | No: Choice of model | Yes: No bot joins. The desktop app captures system audio and the microphone itself, so nothing joins the call. The browser version captures the microphone only. | Partly: Native Mac app. A Mac build of an app that also ships for Windows, the web and mobile. The browser version records the microphone only. | Partly: Local CLI or MCP. Notion has a public API and an MCP server. Hosted, not local. | Yes: No training on your content. Notion states that by default it and its AI subprocessors do not use customer data to train any models, and lists meeting summarisation and transcription among the features in scope. The commitment is made for Notion AI as a whole rather than on the meeting-notes page. | No: Records without you |
| tl;dv | No: On-device transcription | No: Choice of model | Partly: No bot joins. The desktop app records system audio with no bot; a bot joins when the calendar is connected. | Partly: Native Mac app. A Mac build of an app that also ships for Windows and the web. | Partly: Local CLI or MCP. An official API and an official MCP server, both reading tl;dv’s copy rather than one on your machine. Paid plans only; the free plan has no API access. No command line tool. | Yes: No training on your content. tl;dv states on its homepage that it does not use customer data to train its AI models, and that it offers EU data residency. That is marketing copy rather than a policy page. | Yes: Records without you |
| Zoom, Teams and Meet | No: On-device transcription | No: Choice of model | Yes: No bot joins. They are the meeting platform, so nothing has to join. | Partly: Native Mac app. Part of the meeting client itself rather than a separate app, and every one of those clients ships for Windows and the browser as well. | No: Local CLI or MCP | Yes: No training on your content. All three say no. Zoom states it does not use customer content to train its own or third-party models; Microsoft states prompts and responses are not used to train foundation models, naming Teams meetings; Google commits contractually not to train on customer data without permission. Each is a default rather than an absolute. | Partly: Records without you. Teams has a Missed filter for meetings you skipped, and Google Meet sends notes to everyone on the invitation rather than to whoever attended. Zoom’s own pages disagree on whether a non-attendee is sent one automatically, though a host can always share it afterwards. |
| Fathom | No: On-device transcription | No: Choice of model | Partly: No bot joins. Fathom describes its bot-free capture as a beta, on Mac only, and it can only record Zoom, Teams and Meet if you are on the call. | Partly: Native Mac app. A Mac build of an app that also ships for Windows and the web. | Partly: Local CLI or MCP. A REST API and an official hosted MCP server, both reading Fathom’s copy rather than one on your machine. Which plans may call it is not stated. No command line tool. | Partly: No training on your content. Fathom may use de-identified meeting content to train its own models, with an opt-out in account settings. | Yes: Records without you |
| Granola | No: On-device transcription. Granola’s docs state it passes audio from your microphone and system audio to its transcription provider, and does not retain the recording once the transcript exists. | No: Choice of model | Yes: No bot joins | Partly: Native Mac app. A Mac build of an app that also ships for Windows, iPhone, Android and Apple Watch. | Partly: Local CLI or MCP. MCP with per-user OAuth. The API is a paid-tier feature. Neither reads a local copy. | Partly: No training on your content. De-identified data may be used for training, with an opt-out available on every plan and off by default for enterprise workspaces. | No: Records without you |
| Otter.ai | No: On-device transcription | No: Choice of model | Partly: No bot joins. The desktop app supports botless recording for supported meetings; OtterPilot still joins as a participant. | Partly: Native Mac app. A Mac build of an app that also ships for Windows, the web and mobile. | Partly: Local CLI or MCP. MCP with OAuth. The public API is Enterprise only. | No: No training on your content. Otter’s privacy policy, effective 16 June 2026, describes training its models on de-identified audio recordings and on transcriptions, which it notes may contain personal information. | Yes: Records without you |
| Read.ai | No: On-device transcription | No: Choice of model | Partly: No bot joins. The desktop app records without a bot. In bot mode it joins the call, needs host approval, and shows a visible indication in the meeting. | Partly: Native Mac app. A desktop application with separate Apple Silicon and Intel builds, macOS 14 or later, shipped alongside a Windows build. | Partly: Local CLI or MCP. An API and a hosted MCP server, both reading Read.ai’s copy rather than one on your machine. No command line tool. | Partly: No training on your content. The only tool here that asks. Read.ai’s privacy policy makes training opt-in through a named programme you can turn on or off in settings, and excludes data collected through Google Workspace outright. | Yes: Records without you |
The seven requirements
The audio is turned into text on your own machine.
Ask it about the meetings you would not want a stranger to read. Where the audio is transcribed decides whether a recording of your voice, and everyone else’s, exists on someone else’s server at all.
You can choose which model writes the notes, or supply your own API key.
It decides who you are trusting with the transcript, what each meeting costs you, and whether your notes change the day a vendor swaps its model or raises its price.
Nothing is added to the call as a participant.
A bot puts a second copy of the meeting in a second company’s cloud. It is also visible to everyone on the call, which is a drawback or a courtesy depending on the room.
A Mac application built for macOS, rather than one build shared across several platforms.
An app written for the Mac alone reaches the system directly, and its windows, menus and controls are the ones the rest of your machine uses. A cross-platform build targets what every platform it ships on can do, which is the price of running everywhere. Scored on whether the same app also ships for Windows, Linux or the browser, not on a framework nobody publishes.
A command line or an MCP server that reads the copy on your own machine.
This is what turns a pile of meetings into something you can query: a script can pull last week’s calls, and an assistant can answer from a transcript without anything being uploaded first.
Your recordings and transcripts are not used to train models.
Several vendors reserve the right, some by default. The answer is usually in the privacy policy rather than on the pricing page, so it is worth checking before you sign up.
It can capture a meeting you do not attend.
No tool that records your own machine can do this. If a colleague’s call has to be captured while you are elsewhere, you need a bot, and several products here have one.
What each one is for
Listenr
Native Mac app, device capture, on-device transcription
Pick it if you want the recording, the transcript and the notes to stay on your own Mac, and to be able to search them or hand them to Claude without uploading anything. It cannot capture a meeting you are not in, and at $19 a month it costs more than several products here.
- On-device transcription. The default is Apple’s on-device engine, in one language you set. Parakeet, a download, adds 25 European languages with detection. Four of five transcription backends run on the Mac; the fifth, OpenAI, is a cloud option you have to pick.
- Choice of model. A model on your Mac, Apple Intelligence, your own Anthropic or OpenAI key, or Listenr Cloud on Pro.
- Native Mac app. Written in Swift and SwiftUI for macOS and nothing else. There is no Windows, Linux or browser version to keep in step with.
- Local CLI or MCP. The listenr CLI and a local MCP server, free on every plan, reading the database on your own disk. Every command the app has is a CLI verb, and the MCP server has seven read-only tools by default and over 150 when control and settings are enabled by flag.
- No training on your content. With a local model or Apple Intelligence nothing is sent anywhere. Listenr Cloud, the hosted option on Pro, stores neither the transcript nor the notes and keeps only a monthly token count per account.
- Records without you. It records the audio your Mac is playing, so you have to be in the meeting.
Fireflies.ai
Bot and desktop capture, cloud transcription
The one to pick if you want to build on top of your meetings: a GraphQL API on the free tier, over 100 languages in beta besides English, and on Enterprise the option to keep recordings in your own storage bucket. Choose it over a local tool when automation matters more than where the audio is processed.
- On-device transcription. Third-party transcription vendors, under a stated 0-day retention contract.
- No bot joins. The desktop app records system audio with no bot; the bot remains for meetings you do not attend.
- Native Mac app. A Mac build of an app that also ships for Windows, the web and mobile.
- Local CLI or MCP. A GraphQL API available to every user including the free tier, capped at 50 requests a day on Free, plus MCP connectors. Hosted, not local.
- No training on your content. Fireflies states your data is never used to train AI models, internally or externally.
Notion AI meeting notes
Device capture, cloud transcription, inside Notion
If the notes were going to end up in Notion anyway, this is the shortest path there, and it is included with Business. It is also one of the few products that publicly names the services its transcription passes through.
- On-device transcription. Notion names its transcription subprocessors publicly: OpenAI, Anthropic, Fireworks, Baseten, xAI and AssemblyAI.
- No bot joins. The desktop app captures system audio and the microphone itself, so nothing joins the call. The browser version captures the microphone only.
- Native Mac app. A Mac build of an app that also ships for Windows, the web and mobile. The browser version records the microphone only.
- Local CLI or MCP. Notion has a public API and an MCP server. Hosted, not local.
- No training on your content. Notion states that by default it and its AI subprocessors do not use customer data to train any models, and lists meeting summarisation and transcription among the features in scope. The commitment is made for Notion AI as a whole rather than on the meeting-notes page.
tl;dv
Bot and desktop capture, cloud transcription, EU hosted
The one to bring to a European procurement review. EU data residency plus an explicit no-training claim answers the questions a buyer asks, which a file on an employee’s laptop does not, since that raises its own retention and discovery questions.
- No bot joins. The desktop app records system audio with no bot; a bot joins when the calendar is connected.
- Native Mac app. A Mac build of an app that also ships for Windows and the web.
- Local CLI or MCP. An official API and an official MCP server, both reading tl;dv’s copy rather than one on your machine. Paid plans only; the free plan has no API access. No command line tool.
- No training on your content. tl;dv states on its homepage that it does not use customer data to train its AI models, and that it offers EU data residency. That is marketing copy rather than a policy page.
Zoom, Teams and Meet
Built into the meeting platform
Check what your company already pays for before buying anything else. Nobody installs a thing, and speaker attribution is exact because these are the call. The catch is the licence: Teams intelligent recap needs Teams Premium or a Microsoft Copilot licence and an administrator who has turned transcription on, and note taking in Meet needs an eligible Google Workspace edition or a Google AI plan.
- No bot joins. They are the meeting platform, so nothing has to join.
- Native Mac app. Part of the meeting client itself rather than a separate app, and every one of those clients ships for Windows and the browser as well.
- No training on your content. All three say no. Zoom states it does not use customer content to train its own or third-party models; Microsoft states prompts and responses are not used to train foundation models, naming Teams meetings; Google commits contractually not to train on customer data without permission. Each is a default rather than an absolute.
- Records without you. Teams has a Missed filter for meetings you skipped, and Google Meet sends notes to everyone on the invitation rather than to whoever attended. Zoom’s own pages disagree on whether a non-attendee is sent one automatically, though a host can always share it afterwards.
Fathom
Bot and device capture, cloud transcription
The most generous free tier here: unlimited recordings and transcriptions at no cost, with a bot, or without one on a Mac. If price is the constraint, start here.
- No bot joins. Fathom describes its bot-free capture as a beta, on Mac only, and it can only record Zoom, Teams and Meet if you are on the call.
- Native Mac app. A Mac build of an app that also ships for Windows and the web.
- Local CLI or MCP. A REST API and an official hosted MCP server, both reading Fathom’s copy rather than one on your machine. Which plans may call it is not stated. No command line tool.
- No training on your content. Fathom may use de-identified meeting content to train its own models, with an opt-out in account settings.
Granola
Device capture, cloud transcription
The safe default if you do not mind cloud transcription. It runs on more platforms than Listenr, phones included, so your notes follow you off the Mac, and the opt-out from model training is there on the free plan.
- On-device transcription. Granola’s docs state it passes audio from your microphone and system audio to its transcription provider, and does not retain the recording once the transcript exists.
- Native Mac app. A Mac build of an app that also ships for Windows, iPhone, Android and Apple Watch.
- Local CLI or MCP. MCP with per-user OAuth. The API is a paid-tier feature. Neither reads a local copy.
- No training on your content. De-identified data may be used for training, with an opt-out available on every plan and off by default for enterprise workspaces.
Otter.ai
Bot and desktop capture, cloud transcription
The cheapest annual plan here at $8.33 per user per month, and the only one that recognises a returning voice across meetings, which saves real work if you run the same calls with the same people. Read the training clause in its privacy policy before you sign up.
- No bot joins. The desktop app supports botless recording for supported meetings; OtterPilot still joins as a participant.
- Native Mac app. A Mac build of an app that also ships for Windows, the web and mobile.
- Local CLI or MCP. MCP with OAuth. The public API is Enterprise only.
- No training on your content. Otter’s privacy policy, effective 16 June 2026, describes training its models on de-identified audio recordings and on transcriptions, which it notes may contain personal information.
Read.ai
Bot and desktop capture, cloud transcription
The one to choose when the people on the call should see that a recording is happening. In bot mode it announces itself, the host has to admit it, and any participant can remove it by typing “opt out” in the chat, which no silent recorder offers. Its desktop app now records without a bot as well.
- No bot joins. The desktop app records without a bot. In bot mode it joins the call, needs host approval, and shows a visible indication in the meeting.
- Native Mac app. A desktop application with separate Apple Silicon and Intel builds, macOS 14 or later, shipped alongside a Windows build.
- Local CLI or MCP. An API and a hosted MCP server, both reading Read.ai’s copy rather than one on your machine. No command line tool.
- No training on your content. The only tool here that asks. Read.ai’s privacy policy makes training opt-in through a named programme you can turn on or off in settings, and excludes data collected through Google Workspace outright.
Common questions
Which meeting notes app keeps everything on my Mac?
Three do it: Listenr, Anarlog and Meetily all transcribe on the device and keep the library on your disk, and the last two are free and open source. Apple’s own Notes app records and transcribes on a Mac, though Apple does not commit to on-device processing for the summary. Krisp lists on-device transcription as an Enterprise-tier feature, and superwhisper runs its models locally but is mainly a dictation tool. Every other mainstream product sends the audio to a cloud service, which several of them state plainly in their own documentation.
What is the best free AI meeting notes app?
Listenr, if the meetings are personal. Its free plan records and transcribes on your own Mac without a cap, separates the voices on the device, writes notes with a model on the machine or your own API key, and includes the command line tool and local MCP server. The licence is for personal use, so work calls need a paid plan. For commercial use at no cost, Fathom’s free plan is the most generous scored here, with unlimited recordings and transcriptions through a bot, or its bot-free capture on a Mac.
Do I still need a notetaker bot?
Use one if you need meetings recorded while you are elsewhere, or if the people on the call should be able to see that a recording is happening. A tool that captures audio from your own machine cannot do either: it can only record calls you are in, and it is invisible to everyone else. Read.ai’s bot is the clearest example of the alternative, since it requires host approval, shows an indicator, and lets any participant remove it by typing "opt out" in the chat.
Does an on-device model transcribe as accurately as a cloud one?
Not necessarily, and be sceptical of any vendor that says otherwise without publishing a benchmark. Cloud speech recognition generally covers more languages and is often more accurate. You are trading accuracy and language coverage against the audio never leaving your machine, so the answer depends on the meeting: a recruiting call in your own language is a different decision from a multilingual customer workshop.
Score it against your own meetings.
Whether the transcript is good enough for your meetings is not a thing a table can answer. The free plan records, transcribes and writes notes on your Mac, for personal use.