There are only two ways an AI notetaker can get the audio out of your meeting. It either joins the call as a visible participant — a "meeting bot" — or it records the audio already playing on your own machine and never shows up in the participant list. Every "bot-free" notetaker is the second kind. Granola is the best-known one, your video platform's own built-in transcription is the most overlooked one, and a local Whisper setup is the only option that never sends your audio to anybody.
The invisible approach fixes a genuinely awkward problem: nobody wants "Notetaker has joined" appearing in a first sales call. It also swaps that problem for a legal one, which almost nobody selling device-audio capture wants to discuss. We will.
The two architectures
Which bucket a tool sits in tells you almost everything about how it behaves.
Meeting bots (the participant model)
The tool watches your calendar, sees you have a 2pm Zoom, and dials in as its own attendee. It gets a clean, isolated audio stream from the platform, often with video, screen share and speaker labels.
That buys consistency: it works on Zoom, Meet and Teams identically, it works when you are on a plane or not there at all, and attribution is reliable because the platform is telling the bot who is talking. This is the architecture behind most AI meeting tools, and it is why they can promise notes even when you skip the call. It costs you a robot in the room — guests see it, some ask about it, and certain organisations block unknown participants outright, so the bot sits in the waiting room and you get nothing.
Device-audio capture (the invisible model)
The tool runs as a native app on your laptop, taps the system's audio output — everyone else's voices, coming out of your speakers — and separately your microphone. It mixes, transcribes, and never touches the meeting platform.
That buys you invisibility, plus coverage of calls no integration knows about: a browser-tab customer call, a webinar, a conversation over headphones. It costs you everything that depends on being in the room. Skip the call and you get silence. Speaker labelling is a guess based on which channel the audio arrived on. And it gives the other side no visible signal that anything is being recorded — which brings us to the next section.
"No bot" is not "no consent needed"
Recording law does not care about your architecture. In the United States, roughly a dozen states — California, Illinois, Florida, Pennsylvania and Washington among them — require all-party consent to record a private conversation, rather than the one-party consent that governs federal wiretap law and most other states. In the EU and UK, capturing a call means processing personal data, which needs a lawful basis and, in practice, a disclosure. None of these regimes has an exception for "but the recorder was invisible".
The meeting bot, for all its awkwardness, was announcing itself. Take it away and the announcement becomes your job. Apple does not let you dodge this — its built-in phone-call recording plays an audible notice to everyone on the line, by design.
So the honest framing of a bot-free notetaker is not "recording without asking". It is that the disclosure moves from an automated banner to a sentence you say out loud. "I've got an AI taking notes — shout if you'd rather I didn't" costs three seconds and removes the whole issue. Tools implying otherwise are selling you a liability. None of this is legal advice; if you are regulated, get a real opinion.
The bot-free options that actually work
1. Your meeting platform's own transcription
Start here: it is free, already paid for, and everybody forgets it. Zoom, Google Meet and Microsoft Teams all produce transcripts and AI summaries natively — no third party in your data path — and Meet and Teams show a recording indicator, which handles disclosure for you.
Who should use it: anyone whose meetings all happen on one platform, or whose security team would need six weeks to approve a new vendor. Who should not: people spread across three platforms who want one searchable archive. Trade-off: least friction, least intelligence.
2. Granola
The app that defined the category. It runs on your Mac, captures system audio, and — this is the actual idea — merges the transcript with the scrappy notes you typed during the call, producing a summary shaped by what you thought mattered rather than by everything that was said. It has expanded past its Mac-only origins; verify current platform support before committing a team.
Who should use it: people who already take notes in meetings and want them enriched, not replaced. Who should not: anyone needing notes from a meeting they skipped, or clean attribution in a ten-person workshop. Trade-off: the best version of the invisible model, inheriting every one of that model's limits.
3. Apple's built-in recording
Recent macOS and iOS releases record and transcribe audio in Notes, Voice Memos and phone calls, on-device. It is not a notetaker — no action items, no CRM sync — but for a one-to-one where you want a searchable record, it is installed and free.
Who should use it: privacy-first users who need a transcript, not a workflow. Who should not: teams; there is nothing here to share. Trade-off: as private as it gets, as basic as it gets.
4. Superwhisper and the local-transcription family
Superwhisper runs Whisper-class speech models locally on a Mac. Its centre of gravity is dictation, not meetings, and it is worth being straight about that: it is not a drop-in Granola replacement. What it is, is a way to turn audio into text without a byte leaving your laptop — which is why it recurs in discussions of AI tools that work offline.
Who should use it: anyone who needs genuinely local processing. Who should not: anyone expecting summaries and calendar integration. Trade-off: maximum privacy, minimum product. If you can code, the same idea runs free forever: a system-audio loopback (BlackHole on macOS) piped into whisper.cpp, no vendor at all — until an OS update changes audio permissions and it breaks.
5. Descript
A different shape of answer: record the call yourself, import the file afterwards. Descript transcribes it, lets you edit audio by editing text, and takes a file from any source. It is an editing suite that happens to transcribe — but if your meetings become clips or podcasts later, it is the only option here that serves both jobs. The wider set lives in AI transcription tools.
Who should use it: user researchers and content teams who need the audio afterwards. Who should not: anyone who wants notes five minutes after a call ends. Trade-off: a manual step in, far more power out.
Bot vs. bot-free at a glance
| Approach | Shows up in the meeting | Works if you skip the call | Speaker labels | Main catch |
|---|---|---|---|---|
| Meeting bot (Fireflies, Otter, Fathom) | Yes, visibly | Yes | Reliable | Socially loud; blocked by some orgs |
| Granola-style device capture | No | No | Approximate | You must disclose it yourself |
| Platform-native (Zoom/Meet/Teams) | Indicator only | Only if someone records | Reliable | Generic summaries, siloed per platform |
| Local Whisper (Superwhisper, whisper.cpp) | No | No | Weak | Transcript only; you build the rest |
| Post-hoc editor (Descript) | No | No | Decent | Manual import step |
Swipe the table sideways to see every column →
A decision rule you can apply in thirty seconds
- All your meetings on one platform? Turn on its native transcription and stop. You spent nothing.
- Type notes during calls anyway? Granola. Nothing else is built for that behaviour.
- Need notes from meetings you did not attend? You need a bot — there is no invisible way to be in a room you were not in. Our comparison of AI meeting notetakers covers those.
- Audio barred from leaving the building? Local Whisper, and budget engineering time.
- Either way, say the sentence. Every option above is fine once you have told people.
Where bot-free genuinely fails
It fails at scale. A manager who wants every rep's calls transcribed, scored and pushed into a CRM cannot depend on twelve people each remembering to open a laptop app — coverage has to be automatic, and automatic means a bot. It also fails at multi-speaker attribution, which matters enormously in user research and hiring panels. This category is optimised for one person's calls, not an organisation's call data. Buying for a team? Take the bot-based options in our best AI meeting notetakers roundup and accept the robot in the room.
FAQ
Is Granola the only AI notetaker that doesn't join as a bot?
No. It is the most polished, but the same device-audio architecture underpins several newer apps, and your meeting platform's own transcription achieves a bot-free result with no third party at all. Local tools get you a transcript too, just without the summary layer.
Is it legal to record a meeting without a bot announcing it?
The architecture does not change the law. All-party-consent jurisdictions — California, Illinois and several other US states, plus the general expectation under GDPR in Europe — require everyone on the call to be told. A bot-free tool just moves the announcement from the software to you. Say it at the top of the call and you are on solid ground. For anything regulated, take proper legal advice.
Can bot-free notetakers tell who said what?
Only roughly. They separate your microphone from everything coming out of your speakers, so "you vs. them" is clean. Splitting three voices out of a single mixed far-end stream is much harder — the clearest quality gap versus bot-based tools, which receive labelled streams straight from the platform.
Do these tools work on Windows?
Increasingly, yes — but the category grew up on macOS and support genuinely varies. Check the download page rather than trusting any article, including this one. On Windows, the platform-native route is usually the pragmatic answer.
Where does the audio actually go?
That depends on the tool, not the architecture. Most bot-free notetakers still upload audio to their servers to transcribe it; "no bot" says nothing about "no cloud". If keeping audio on your machine matters, you need a tool that explicitly says the speech model runs locally — verify that in the documentation, not the homepage.
Where to go next
For the full field, bots included, browse AI meeting tools, or narrow to raw speech-to-text in AI transcription tools. The free tools list shows what a category costs before you commit. Running a bot-free notetaker we have missed? Submit it — this corner of the market moves faster than the search results do.