Google’s new Mac app writes up your meetings without sending anything to the cloud

Google has released Google AI Edge Foresight, an experimental Mac app that sits in on your meetings, listens in the background and turns the scrappy bullet points you type into full notes, with every step running on the computer itself rather than in Google’s cloud.

The app launched on October 6 alongside EmbeddingGemma 2, a new open model from Google DeepMind, and it is Google’s showcase for what that model can do. It puts Google up against meeting note-takers such as Granola, with one big difference: Google says nothing you record or import leaves your Mac.

How Foresight works

Foresight captures your microphone and your Mac’s system audio, so it works with Google Meet, Zoom or any other call, and with face-to-face meetings too. Google’s pitch is that you stop taking detailed notes: you “scribble shorthand bullets for things you want to remember” and Foresight expands them into polished notes using the live transcript. A line such as “transcribes audio” becomes a full sentence explaining the point, pulled from what was actually said.

It can also learn about your work. You point it at project folders, reference material, diagrams, your calendar and Google Drive, and it indexes them without you having to organise anything. Google lists support for PDFs, Google Docs, Microsoft Office files, plain text, Markdown and web bookmarks, with custom labels to sort them.

That library then feeds two features:

• Ask: a chat that answers questions and summarises from your own files.
• Live Assistance: while a meeting is running, Foresight listens for questions or topics that match your files and shows “proactive context cards with exact citations”.

It is a similar idea to Google Cloud’s new Gemini agent, which also works across your mail, calendar and Drive, except that one runs in Google’s data centres and Foresight runs on your laptop.

What runs on your Mac

Foresight pairs two Google models. EmbeddingGemma 2 is a 740 million parameter embedding model that maps text, images, video frames and audio into a single “vector space”, so one model can match a typed question against a document, a diagram or a stretch of a recording. Gemma 4 does the writing. Together they run what Google calls a “fully on-device RAG pipeline”, meaning the app looks up your own material and grounds its answers in it without a network connection.

Google says “all AI inference runs 100% locally” and that no files, recordings or transcripts are ever sent to its servers. Because nothing depends on the internet, it says Foresight works on a plane or in an air-gapped secure building. There is no subscription: Google points to “no cloud subscription costs” as one of the benefits of running locally.

EmbeddingGemma 2 itself is open, released under the Apache 2.0 licence on Hugging Face and Kaggle. Google says the text-only part needs about 191MB of active memory and the full multimodal model about 567MB on a Pixel 11 Pro, and it will come to Android apps through ML Kit “in the coming weeks”. Two other demos, Instant Media Search and Video Moments Finder, are in Google’s AI Edge Gallery app for Android and iOS. If you want to run models like this yourself, our guide to running local AI models on a Mac mini covers the hardware.

Requirements and where you can get it

Foresight is a direct download from Google’s Foresight page, a disk image of about 160MB, and it fetches its on-device models the first time it runs. Google says it is “optimized for Apple Silicon” Macs. It has not said whether older Intel Macs are supported, and there is no Windows or Linux version.

Google has not listed any country restrictions, and the download does not go through the Mac App Store. It also hasn’t said which languages the transcription handles; its product page and demos are in English. Foresight is labelled experimental, so expect rough edges, and it joins two other Google AI Edge apps for the Mac, AI Edge Gallery and AI Edge Eloquent, an offline transcription tool.

One thing worth knowing before you use any meeting recorder: the rules on recording other people differ between countries and US states, and some require everyone on a call to agree. Google’s page doesn’t say how Foresight tells other participants that it is listening.

Why it matters

Most AI note-takers send your meetings to a server to be transcribed and summarised, something many companies restrict for confidential meetings. Foresight is Google showing that small open models are now good enough to do that job on a laptop, and if it works as promised, the privacy argument for cloud note-takers gets much harder to make. Microsoft is pushing the same way with Copilot working on the files on your PC.

Sources: Google Developers Blog, Google AI Edge Foresight, Google: EmbeddingGemma 2, EmbeddingGemma 2 on Hugging Face.