Google AI Edge Foresight: Offline Meeting Notes on Mac

The short answer: Google AI Edge Foresight is an experimental Mac app (Apple Silicon) that records system audio and the microphone, transcribes and writes notes locally with Gemma models, and works offline for capture and notes. It is a strong signal that meeting AI can run on your own device, but it is not yet a product you can roll out to a team. If you need on-device meeting AI across Mac, Windows, iPhone and Android today, Hedy already ships it.

On 6 October 2026, Google published a developer blog post about EmbeddingGemma 2, its new small embedding model. It also introduced an app called Google AI Edge Foresight, which is the most interesting thing to happen to on-device meeting AI this year.

In Google’s words, Foresight is “an experimental app” that brings “a context-aware meeting companion directly to your device.” It works with any meeting platform, “even completely offline,” and is “powered by EmbeddingGemma 2 and Gemma 4 models.” Google lists “no cloud subscription costs” among the on-device benefits. It expands your shorthand notes from the live transcript, the way Granola does, and it can search your own files for context. Coverage from 9to5Google and others adds that it is a Mac app and that transcripts and audio are processed on the device.

Why this matters

For two years, the standard answer to “why does my notetaker need the cloud?” was that good transcription and good summaries need big models on big servers. Vendors said it, and most buyers accepted it.

Foresight is Google, the company with arguably the largest cloud AI business in the world, shipping a meeting notetaker whose headline feature is that it does not need one. That does not prove local models are as good as cloud models for every task. It does settle the narrower question that matters for this category: a laptop can now transcribe a meeting and write useful notes from it without sending audio anywhere.

We have argued that this is where the category is heading, in our guide to what on-device AI actually means. The argument is about the copy. A vendor that never receives your audio cannot leak it, train on it, or hand it over under subpoena, and has a much harder time being accused of eavesdropping on it, which is the theory a court let proceed against Otter in the Otter.ai lawsuit. Encryption protects a copy. On-device processing means the copy is never made.

What Foresight is not, yet

Read Google’s wording carefully, because it is careful.

It is an experiment. Google calls it one. Experimental apps can change, stall, or disappear. That is a reason not to build a team workflow on it today.

It is Mac only. It is built for Apple Silicon Macs. Google’s Android and iOS offerings are separate demos, so there is nothing for the in-person meeting where the only device in the room is a phone.

“Offline” has limits. Google says the app works completely offline, but connecting Google Drive files needs a network. Treat the offline claim as covering capture, transcription and notes.

There is no stated consent feature. Like every bot-free tool, Foresight captures system audio silently. Nothing in Google’s announcement mentions a recording notice, a disclosure prompt, or guidance for all-party-consent states. That is not unusual. It is still the open legal question for the whole bot-free category, and the Granola class action is testing it now.

Nobody has audited it. “Works offline” is a claim we can test, and we will. Until then it is Google’s description. Our airplane test is simple: turn off the network, record a meeting, and see whether you still get a transcript and a summary.

How it compares with the on-device tools we already rank

Foresight joins a short list. Among the 28 tools in our rankings, two stand out on the airplane test:

• MacWhisper runs entirely on the Mac with no account. It transcribes well and detects meetings automatically, but its AI summaries stay offline only if you install and configure a local model yourself, and it has no cross-meeting memory or team features.
• Hedy runs speech recognition on your device, and since its May 2026 Local AI release it can run the rest of the pipeline (summaries, detailed notes, chat and live coaching) on-device on supported hardware. That covers Apple Silicon Macs, capable Windows GPUs, iPhone 15 Pro and newer, M-series and A17 Pro iPads, and, since September 2026, Android phones with a supported GPU and about 8 GB of memory. It is a shipping product with SOC 2 Type I (completed April 2026), a HIPAA BAA on its Enterprise plan, and contracts that bar its AI providers from training on anything that does touch its cloud.

Foresight sits between them: more meeting-aware than MacWhisper, much narrower than Hedy, and not yet a product. Its file-search feature is a genuinely good idea. Bringing your own documents into the live meeting context is something most cloud tools still do badly.

Several popular tools describe themselves as “on-device” because the audio capture happens on your computer, while transcription and summaries still run on their servers. Foresight’s arrival makes that distinction easier to explain: if Google can do the whole thing locally on a Mac, ask any vendor using the word why they cannot.

What to do with this

If you are on a Mac and curious, try Foresight on a meeting where you would not mind losing the notes. Run the airplane test. Tell the other participants you are recording, because the app will not. The same goes for Hedy: bot-free capture, on-device or not, never replaces consent.

If you need something you can deploy to a team today, across Windows and phones as well as Macs, with compliance paperwork behind it, Foresight is not that yet. Start with our offline guide, the Mac guide, our privacy rankings and the fact sheet, which lists where every tool we track actually processes audio.

We will review Foresight properly if Google turns it into a product. In the meantime, the largest cloud company in the world has just agreed that meeting audio does not need to leave the device.

Sources: Google Developers Blog, “Bring multimodal semantic search to the edge with EmbeddingGemma 2” (6 October 2026); 9to5Google, 8 October 2026. Details as of 27 October 2026.