Google AI Edge Foresight Brings Offline Meeting Notes to Mac

A Google notetaker that keeps your meetings off its servers

Google has released a meeting assistant that does the one thing its own cloud tools are built not to do: it keeps your data on your device. The app, Google AI Edge Foresight, launched on October 6, 2026, as a free, experimental macOS download. It arrived alongside a new open model called EmbeddingGemma 2.5 Foresight records meeting audio, transcribes it, turns your rough notes into finished ones, and answers questions about what was said. Google says it does all of this on the Mac itself, so files, meeting audio and notes "never leave your computer."2

The coverage agrees on the basics. CNET calls it a notetaker that transcribes "without sending your conversations to a data center" and points out that it is Mac-only for now.1 The Verge, following TechCrunch's earlier report, says it can transcribe meetings and audio files completely offline.2 Several outlets also describe it as a direct, local-first challenger to Granola, the AI notepad that has become popular with knowledge workers.313

The launch matters less for any single feature than for what it says about where AI productivity tools are going. A big cloud company has shipped a showcase app built on the idea that the cloud is optional.

How Foresight works

The design will be familiar to anyone who has used Granola. Foresight opens a split window. You type shorthand on the left while a live transcript runs on the right, captured from the Mac's microphone and system audio.11 Because it listens at the operating-system level, it works with Google Meet, Zoom and in-person conversations. It doesn't need a plug-in for each platform, and it doesn't send a visible bot into the call.5 After the meeting, it expands your bullet points into polished notes using the transcript as context.1

This is where it starts to look like a research copilot as well as a transcriber. You can build a "knowledge source" from PDFs, Google Docs, Microsoft Office files, plain text, Markdown and web bookmarks before a meeting starts.11 An Ask window lets you query that material. A feature called Live Assistance listens during the call and answers questions as they come up, though How-To Geek notes that Google's product page gives few details about how it works.11 According to How-To Geek, Foresight also builds a knowledge graph that links meetings, your notes and attached files, so you can go back to that context later.11 Labels, filters and search all work without a network connection, and Google says the background processing uses little battery.1

Based on these descriptions, Foresight combines three product categories that have usually been sold separately: a meeting assistant, a note-taking app, and a document-grounded Q&A tool.

The model story is messier than first reported

The biggest disagreement in the coverage is about which models run the app. The Verge's report says Foresight runs on EmbeddingGemma 2.2 Firstpost describes "an on-device Gemma 2 model with 740 million parameters," which appears to blur two different model names together.7 How-To Geek and NewsBytes give a fuller picture. EmbeddingGemma 2 is a 740-million-parameter multimodal embedding model, and the conversational assistant runs on Gemma 4.1113

Progressive Robot argues that describing the app as EmbeddingGemma 2 alone is incomplete. It cites Google's developer materials, which pair EmbeddingGemma 2 for finding local files with Gemma 4 for reasoning.5 That reading makes sense. An embedding model maps text, audio, images and video into a shared vector space so related material can be found later, but it doesn't write prose.7 One gap remains open: Google has not said which model produces the transcript, so claims about speech recognition are still educated guesses.5

The distinction matters for buyers. Foresight's quality depends on how well a small local model can reason over retrieved context, which is a different challenge from accurate transcription.

Why local-first is the real story

Cloud-based notetakers have improved quickly, but they all come with the same trade-off. CNET notes that most popular tools, including Scribe, Granola and Notion AI, store and process data in the cloud, which creates privacy exposure. It names Meetily and Whisper-based tools as the existing local alternatives.1 Commentary on Winzheng explains why this matters for meetings in particular. Meetings often cover trade secrets, roadmaps, personnel issues and legal matters, so compliance teams tend to resist sending that audio to someone else's servers.4 The same analysis notes two other benefits of running locally: lower latency and no per-use API cost.4

Google's FAQ, as quoted by Progressive Robot, says the app works on planes and in air-gapped secure facilities.5 For lawyers, consultants, clinicians and others bound by confidentiality rules, that is the main selling point.

Price is part of the appeal too. Google advertises "no cloud subscription costs."6 By comparison, Granola charges $14 per user per month for Business and $35 for Enterprise.5 One Google Cloud developer advocate called it a free notetaker in which every premium feature works without a subscription.8 That post is enthusiasm from inside Google, not independent testing, but it shows how Google wants the app to be seen.

Google is competing with itself

The odd part of this launch is how different Foresight is from the rest of Google's note-taking products. In Meet, Gemini's "Take notes for me" saves notes to the organizer's Google Drive and emails a link afterward. Since April 2026, it has also worked as a capture tool for in-person and other-platform meetings, still running on Google's cloud.12 Google Voice now offers Gemini call notes as well, with the recording and transcript delivered by email or through call logs.16 In September, Google added a mobile audio recorder and real-time voice conversations to Gemini Notebook, the product formerly known as NotebookLM.18

Foot reviews of the cloud tools have been mixed. One tester found that Meet's built-in notes were less detailed than those from dedicated tools such as Granola or Fellow, and estimated transcription accuracy at 80 to 85 percent.17 Meanwhile, Meet has started flagging third-party bots from services such as Otter, Fireflies and Fathom as security risks, while Gemini's own notes are unaffected.12

So why release a local app at all? The most likely answer comes from where the product sits. Foresight comes from Google AI Edge, the company's group for on-device AI technology, not from the Workspace team.11 NewsBytes suggests the release may partly be a way to show off what the Gemma family of offline models can do.13 Firstpost and NewsBytes both report that Google may also be preparing a consumer notetaker that would work across video-calling apps through Gemini, using cloud models.1314 On that reading, Foresight looks less like a reversal of Google's cloud strategy and more like a working demo for developers that also tests demand for private AI tools.

The limits are real

The current release is clearly narrow. Foresight only runs on Apple Silicon Macs. Intel Macs aren't supported, and it's a roughly 160 MB direct download rather than a Mac App Store listing.5 CNET notes that there is no word on a Windows version or Gmail integration.1 Welcome.ai argues that launching only on macOS limits its early reach and could leave room for competitors.9

It also has no team features. Progressive Robot points out that Google has published nothing about sharing, team workspaces or admin controls. Granola, by contrast, offers SSO, admin controls, mobile apps and integrations with Notion, HubSpot and Slack.5

The privacy promise also has limits that several outlets point out. Progressive Robot notes that keeping data local moves the risk instead of removing it. A transcript stored on a laptop is only as safe as that laptop's disk encryption and screen lock. And the Google Drive connection means not every knowledge source is truly offline.5 Tech Journal recommends testing the app on low-risk meetings first, limiting its file permissions, and getting participants' consent. Processing audio locally does not change recording law.6 Because there's no visible bot, other people in the call may not know they are being transcribed, so users need to tell them.5

Verdict

Foresight is not a finished product, and Google isn't presenting it as one. Still, it shows that the core features of paid AI meeting assistants can run on a laptop at no cost: live transcription, note enhancement, and question answering grounded in your own documents. Once Google offers that for free, cloud-only competitors will have a harder time justifying a subscription on features alone, and privacy-first startups lose some of their uniqueness.

The next phase of competition will likely be about trust, platform coverage and team features, not transcription accuracy. Winzheng frames it as a question of who best understands meeting context while protecting privacy.4 Google has staked out a position on privacy. Whether it builds on Foresight or folds what it learns back into cloud-based Gemini will show how seriously it takes that position.