The phrase "AI meeting copilot" gets thrown around a lot. Half the apps that sell it send a bot into your Zoom call, capture the audio on someone else's server, then charge you per minute to read it back. That's not what this article is about. This is about a meeting copilot that lives on your Mac, doesn't join the call, doesn't upload the audio, and turns the transcript into structured notes after the meeting — without anyone else's infrastructure seeing your conversation.

What does "on-device AI meeting copilot for Mac" actually mean?

Three words, three honest definitions. On-device means the speech recognition runs on the Neural Engine of your Apple Silicon Mac. Audio is captured and processed locally. In MetaWhisp's local mode, the audio file never leaves the laptop — there's no upload step in the pipeline at all. No round trip to a server, no API key required for transcription, no waiting for someone else's GPU. AI meeting copilot means a tool that handles the boring part of meetings for you — the capturing, transcribing, note-taking, action-item extraction — so you can stay present in the conversation instead of typing. The "AI" part doesn't have to mean a cloud LLM. It can mean a local model, a BYOK model, or no model at all if you only need the transcript. For Mac means macOS 14 or later on Apple Silicon (M1 or later). Whisper large-v3-turbo runs comfortably in WhisperKit on the Neural Engine, which is the technical reason MetaWhisp exists as a Mac app and not, say, a web service. On Intel Macs, Whisper doesn't hit usable speeds. On Apple Silicon, it's fast enough to transcribe a meeting in roughly the time the meeting took. Put together: a Mac app that runs Whisper locally, captures your mic and the meeting app's audio, transcribes both, and gives you an AI-shaped way to turn the transcript into something useful when the call ends.

How does MetaWhisp actually capture a meeting (the two-track part)?

The single most important thing to understand about MetaWhisp as a meeting tool is this: when you start a recording, it captures your microphone audio and your Mac's system audio as two separate tracks, not one mixed file. Why this matters:

• Your voice stays in its own track. The other participants' voices — which arrive through Zoom, Google Meet, Microsoft Teams, Webex, Slack huddles, anything — stay in a different track.
• Later, you can process them separately, mute one, or run AI on just one.
• If something sensitive happens on the other side (a customer's name, a number you shouldn't be storing) you can trim only that track.
• You can ship just the mic track to your own ASR for training data without leaking everyone else's voice.

Here is the part I have to be careful about: two tracks is not the same as speaker diarization. Speaker diarization is the technical capability to label "Person A said this, Person B said that" — figuring out who is who. MetaWhisp does not do that. The two tracks tell you "your mic vs. everyone else," nothing more. If you need real "who said what" labels, that's a feature MetaWhisp does not ship today and I won't pretend otherwise. Recording itself is off by default. You have to turn it on, every time, deliberately.

Live transcription vs. live AI coaching — what does MetaWhisp actually do during a call?

A lot of AI meeting tools promise a live copilot experience: a panel pops up with "questions to ask" or "objection handling" prompts as the call happens. Otter, Read AI, and a dozen others sell exactly that. Honest version: MetaWhisp does not ship a live AI coach overlay that whispers prompts in your ear during a call. What it does do:

• Live transcription in the MetaWhisp app window. As your mic and the meeting audio come in, Whisper large-v3-turbo produces a transcript that updates in near real time. You can see what's being said — useful for accessibility, useful for staying on track.
• Recording of both tracks for later review, search, and AI processing.
• Post-call AI: after you stop recording, you can run Structured, Correct, or Rewrite on the transcript to get notes, action items, cleaned-up prose. The AI only ever sees the text, not the audio.

The value is the same as the live-coach apps for many meetings: by the time the call ends, you have a transcript, a recording, and a set of structured notes waiting in your clipboard or your notes app. The difference is timing — MetaWhisp's AI work happens after the call, not during it. For sensitive conversations, that's often a feature, not a bug.

Post-call AI: where the copilot becomes real work

If live transcription is the appetizer, post-call AI is the actual meal. MetaWhisp offers three post-processing modes that work on the transcript text:

• Structured: turns a raw transcript into organized notes with sections like Summary, Decisions, Action Items, Open Questions. This is the closest thing to the "AI summary" you get from Otter or Fireflies, and it runs locally if you point it at a local LLM, or via your own API key, or via Pro's built-in cloud AI.
• Correct: cleans up filler words ("um", "uh", repeated phrases), fixes obvious transcription errors, gives you readable prose. Good for transcripts you plan to paste into a doc without re-reading.
• Rewrite: rewrites the transcript in a chosen voice or style — meeting minutes, blog draft, follow-up email. Less useful for raw meetings, more useful for client calls where you'll quote the substance later.

The full breakdown lives on the processing modes page, but the meeting-copilot pattern most people settle into is: record → transcribe locally → run Structured → paste the result into your notes tool or Notion or Obsidian or Todoist. A real-world flow from my own week: a 45-minute client call ends. I hit Stop. MetaWhisp finishes the local transcript in maybe 20 seconds. I click Structured with my BYOK OpenAI key. I get back a doc with four action items, three open questions, and a 200-word summary. I paste it into my CRM, send a recap email to the client, and I'm done before the kettle boils.

Consent, recording laws, and "tell your participants"

This is the part that has to come before anything else, even if it slows down the article. Recording a conversation without the other person's knowledge is illegal in many jurisdictions. In the US, the all-party-consent states (California, Florida, Illinois, Maryland, Massachusetts, New Hampshire, Pennsylvania, Washington, and others per state-by-state surveys of recording law) require explicit consent from every participant. In the EU, GDPR imposes strict rules on processing personal data, including voice, that apply whether the recording lives on your laptop or someone else's cloud. The specifics vary by jurisdiction. The principle doesn't. MetaWhisp can't tell you whether you're allowed to record a meeting. That is your legal responsibility, not the app's, and no tool on earth can grant you that permission. What MetaWhisp does:

• Recording is off by default. You turn it on, deliberately, every time.
• Nothing is uploaded unless you explicitly turn on a cloud feature.
• The captured audio stays on your Mac in local mode.

What you do:

• Tell people you're recording. "I'm going to record this so I can take better notes — okay?" is enough in most one-party-consent jurisdictions.
• In all-party-consent places, get explicit verbal or written confirmation before recording.
• For work meetings, check your company's recording policy and your employment contract.
• For client calls, send a calendar note the day before: "I'll be recording for note-taking. Let me know if you'd prefer I don't."

I run MetaWhisp every day. The two-second verbal consent at the start of a call has never once made a meeting awkward — and it's saved me from the kind of misunderstanding that ruins client relationships.

When does MetaWhisp beat a meeting bot (and when doesn't it)?

I'll be honest about both.

Dimension	MetaWhisp (on-device)	Meeting bot (Otter, Read AI, Fireflies)
Joins the call	No — captures audio on your Mac	Yes — appears as a participant
Audio storage	Local file on your disk by default	Vendor cloud (per their privacy policy)
Live AI prompts	No — post-call only	Yes, on most paid tiers
Speaker labels	No (no diarization)	Yes, on most paid tiers
Cost model	Free local; Pro $30/yr adds cloud AI	Typically per-seat, per-month (per their pricing pages)
Setup	Install Mac app, hit record	Connect calendar, bot joins automatically
Internet required	Only for cloud AI / cloud transcription	Yes

MetaWhisp wins when:

• You don't want a third party on the call (client work, therapy, legal, healthcare).
• The conversation contains personal, financial, or competitive information.
• You're on a platform that blocks bots, or with people who don't like them.
• You want the audio to stay on your machine, full stop.

Meeting bots win when:

• You want "who said what" automatically labeled.
• You want live AI prompts during the call.
• You're on a team that lives inside the bot's workspace (searchable history, shared notes).

If you want a deeper breakdown, I wrote it in how to transcribe a meeting without a bot joining. The short version: for solo work, sensitive calls, and privacy-sensitive practices, on-device wins on cost, control, and trust.

What does MetaWhisp NOT do for meetings (so you know the limits)?

I promised to be honest, so here is the honest list:

• No speaker diarization. MetaWhisp cannot tell you which participant said which line. The two-track recording only separates "your voice" from "everyone else's."
• No live AI coaching. No "questions to ask" panel, no objection-handling prompts, no real-time summarization of what was just said. AI work happens after the call.
• No automatic meeting joining. MetaWhisp doesn't read your calendar or pop into a Zoom link. You launch the call yourself; MetaWhisp just listens.
• No iOS app yet. It's on the roadmap for later in 2026. For now: macOS only.
• No semantic transcript search. You can search inside one transcript. Cross-meeting semantic search ("every meeting where we discussed churn") isn't shipped.
• No domain-specific accuracy claims. Legal, medical, and other specialized vocabulary haven't been benchmarked. Our own run on LibriSpeech test-clean came out at 2.76% WER (~97% accuracy) — but that's a clean-readings benchmark, not a courtroom deposition.

If any of those are dealbreakers for your use case, you should know now, before you set up the workflow. I'd rather lose a reader to an honest limitation than win a customer on a feature that doesn't exist.

Free tier vs. Pro for the meeting copilot flow

You can do a lot without paying MetaWhisp a cent. Here's the matrix:

Feature	Free + local	Free + BYOK	Pro ($30/yr or $7.77/mo)
Two-track meeting recording	Yes, unlimited	Yes, unlimited	Yes, unlimited
Local Whisper transcription (99 languages)	Yes, unlimited	Yes, unlimited	Yes, unlimited
Structured / Correct / Rewrite	No	Yes, via your OpenAI or Cerebras key	Yes, built-in cloud AI
Cloud transcription (Whisper large-v3-turbo)	No	No	Yes, with daily cap
Audio leaves your Mac	Never	Only transcript text → your own API	Yes, for cloud features (audio + text)
Priority support	No	No	Yes

BYOK = "bring your own key." You paste in your OpenAI or Cerebras API key, and MetaWhisp sends the transcript text (not audio) to your account at your provider. You pay OpenAI or Cerebras directly, not MetaWhisp. The audio never touches MetaWhisp's servers because there are no MetaWhisp servers in local mode. Pro is the path if you'd rather not manage an API key, or you also want MetaWhisp's built-in cloud transcription for situations where your Mac isn't around. Pro features do send data to servers — say so if you're handling anything sensitive. The full breakdown is on the pricing page.

How do you set it up in 10 minutes?

Quick recipe. The full step-by-step with screenshots is in how to transcribe meeting minutes on a Mac; this is the short version.

1. Install MetaWhisp. Free download, macOS 14+, Apple Silicon. ~950 MB model download the first time you run it. No account, no email, no signup form.
2. Grant mic permission and system-audio permission when macOS prompts. Say yes to both. Without system audio permission, the second track is silence — which is the #1 reason people think MetaWhisp "isn't working."
3. Open Preferences → Recording Sources and enable both your microphone input and the system audio capture. Both tracks should appear in the input meter.
4. Open your meeting app (Zoom, Meet, Teams, Webex, FaceTime, anything). Start or join the call.
5. In MetaWhisp, hit Record. Both tracks start capturing. You'll see the live transcript begin updating in the app window.
6. During the call: stay present. Don't transcribe by hand — that's the whole point. If you need to mark a moment, jot a one-word note in your notes app; the recording captures everything.
7. End the call. Hit Stop in MetaWhisp. The two WAV files and the merged transcript are saved on your disk.
8. Optional: run AI processing. With a BYOK key or Pro, click Structured on the transcript. Paste the result into wherever your notes live.

Ten minutes, one-time setup. After the first recording, step 3 onward is muscle memory.

Pro tip: Before any high-stakes client call, do a 30-second test recording with both tracks enabled, then play the two WAV files back. Verifying that you actually captured the system audio (and not silence) is the single biggest failure mode I see people hit. It's almost always a permissions issue, not a bug. The second most common is forgetting to grant Screen Recording permission to the meeting app itself.

Frequently asked questions

Related reading

• Meeting transcription without a bot joining your call — the longer version of the bot-vs-local comparison
• How to transcribe meeting minutes on a Mac — the step-by-step setup guide with screenshots
• Processing modes: Structured, Correct, Rewrite — the full breakdown of what AI post-processing actually does
• Pricing: Free local vs Pro cloud — the comparison if you want built-in cloud AI without a BYOK key