Meeting notes that never leave your Mac
WalkAITalkie transcribes meetings fully on-device: minutes, action items and an email draft are built locally on your Mac, no cloud, unlike Otter or Granola.
TL;DR
- WalkAITalkie is getting a Meeting Mode: recording, transcription and summarization all run on your Mac.
- No cloud upload, no server-side processing like Otter or Granola.
- On-device transcription (Parakeet/WhisperKit + MLX) produces minutes, structured notes and action items.
- Speaker diarization separates who said what, and an optional mailto draft closes out the workflow.
- Ships first in the direct/Sparkle beta channel. App Store submission is a later phase.
The privacy problem with cloud notetakers
Otter and Granola solve a real problem. You sit in a meeting, you want to listen instead of scribbling notes, and at the end you still get a usable summary. That works well, and I understand why so many teams use these tools every day.
But the mechanism behind it is always the same: the audio leaves your machine. It gets uploaded to a server, transcribed there, summarized there, and stored there. Between "I am speaking" and "I have minutes" sits someone else's data center.
For a lot of meetings that is fine. For others it is not. A client call with contract details, an internal roadmap discussion, a conversation with legal or personnel sensitivity: once that content lands on someone else's infrastructure, confidentiality depends on a privacy policy, a retention window and a security practice you do not control. Most vendors are careful about this. But "careful" is not the same as "not possible in the first place".
That is the point where I start wanting an on-device alternative. Not because cloud processing is inherently unsafe, but because for sensitive content the only guarantee that actually matters is: the data never left the device. There is nothing to leak if it never goes anywhere else.
There is a second, more practical concern too. Many cloud notetakers depend on an ongoing subscription, and what happens to stored transcripts once you cancel, or once the vendor shuts the service down, is often unclear. Some vendors fold snippets from meetings into training data unless you actively opt out, disclosed somewhere in the terms of service that almost nobody reads before their first meeting. What is left is a dependency on someone else's infrastructure, which is exactly what you do not want for sensitive conversations.
Meeting Mode: recording, transcription, summary, all local
This is where the new Meeting Mode in WalkAITalkie comes in. WalkAITalkie is my macOS menu-bar app for push-to-talk dictation, built in Swift with WhisperKit and MLX, live on the Mac App Store. Meeting Mode extends the same principle from dictation to in-person meetings.
The flow is deliberately simple. You start the recording from the menu bar, the same way you would for regular dictation. Your Mac's microphone picks up the conversation in the room, no separate meeting tool, no bot joining a video call. After the meeting, transcription runs directly on the device, via Parakeet or WhisperKit combined with MLX for on-device inference on Apple Silicon.
From the raw transcript comes a structured summary: minutes of what was discussed, clearly organized notes, and a list of action items that came out of the conversation. Not an unstructured block of text you still have to sort yourself, but a result you can act on directly. The entire path from recording to finished summary runs on the Mac. There is no step in between where audio or text gets sent to a server.
In practice that means no network request during processing, no wait tied to an API response, and no per-minute transcription bill. The compute load sits on the Neural Engine and GPU of your own Mac, not on rented server capacity. That also makes the flow usable when a stable internet connection is not available, on the road or in a meeting room without reliable WiFi. For a tool meant to work exactly at the moment a meeting is happening, that is not a side effect. It is part of the design.
Speaker diarization and the mailto draft as the finishing step
Minutes without recognizable speakers are only half useful. Who made the proposal, who agreed to which follow-up? That is why speaker diarization runs alongside the transcription: the system distinguishes which passages came from which person in the room and attributes statements accordingly. That makes the minutes readable, not just complete.
At the end of the workflow sits an optional email draft. WalkAITalkie opens a mailto link that launches your default mail client with a pre-filled draft, subject, summary, action items. Nothing is sent automatically, no cloud email API running in the background. You see the draft, review it, adjust whatever is needed, and send it through your own mail app. The follow-up step that so often gets skipped after a meeting is done in seconds, without any intermediate step giving up your control over sending it.
That choice to stay with mailto instead of a direct send integration follows the same principle as the transcription itself. A direct send through a cloud API would mean storing credentials to your mail account somewhere, and routing the finished text through yet another service before it gets sent. The mailto detour is marginally less convenient, one extra click, but it keeps control over the final word entirely with you. For an internal recap that might not matter much. For a summary with confidential details, it is the difference between "I sent it" and "it was sent for me".
The format library: one model, many output shapes
What sits behind Meeting Mode is not limited to meetings. The underlying engine is an on-device pipeline of recording, transcription and structured summarization. The meeting minutes are one output format of that pipeline, but the architecture generalizes into a format library.
Concretely: the same model that turns a meeting into minutes can turn a different recording into a different structure. One example that convinced me is a doctor's visit. You record the conversation, and instead of a generic transcript, you get a structure that makes sense for that context, with the relevant points organized rather than left as running text. The underlying logic stays the same: recording in, matching structure out. Which format makes sense depends on the use case, not on the underlying technology.
That generalization is the part that makes WalkAITalkie more interesting to me than a pure meeting tool. It is not a single feature, it is a reusable pattern: on-device audio processing with an interchangeable output structure. In practice, that also means new formats do not need new core infrastructure. The recording and transcription pipeline stays the same, only the summary template changes depending on the context you are capturing. That is the difference between a tool that handles one function and a tool that solves a recurring problem, namely turning spoken content into a structure that fits its purpose, without the recording ever leaving the device.
Honest status: the direct channel comes first
I do not want to pretend Meeting Mode will be available to everyone in the Mac App Store tomorrow. It ships first in the direct channel, through the Sparkle update mechanism that WalkAITalkie uses for beta versions. If you get the app directly from the website, you get new features earlier, before they go through the App Store review process.
That is not an excuse, it is a deliberate order. A feature like Meeting Mode benefits from being tested first with real meetings under real conditions, before it ships in the App Store version, which reaches far more users and needs to be correspondingly stable there. App Store submission is a later phase in the rollout, not a promise pushed indefinitely. Whoever is on the direct beta channel sees the feature first. Everyone else gets it once it clears review.
Why "never leaves your Mac" is the one axis that counts
You can compare meeting-notes tools on plenty of axes: transcription quality, summary style, integration with calendars and note apps, price. On most of those axes, there are good cloud solutions, and Granola or Otter do things today that WalkAITalkie does not yet do.
But there is one axis where a cloud notetaker cannot structurally match an on-device tool: whether the data ever leaves the device at all. With cloud processing, the answer is always "yes, at least briefly, on the way to processing". With a local path, the answer is a clear "no". That is not a gradual improvement in security. It is a different security model altogether.
As a solo builder who runs my own client calls and depends on confidentiality there, that is exactly why I am building this direction into WalkAITalkie. Not because cloud tools are bad, but because for certain conversations there is no substitute for "no server ever listened in". That is the one axis that counts when the content of a meeting genuinely needs to stay confidential.
Conclusion
Meeting Mode is not an attempt to beat Otter or Granola on every category. It is a different principle: recording, transcription and summarization stay on the Mac, from the first second to the finished email draft, shipping first in the direct channel, with the format library as the bigger picture behind it.
If you are curious, you can find WalkAITalkie at walkaitalkie.com. And if you want to know more about how I build on-device tools into real workflows beyond meeting notes: Contact is the direct route.