Tools AI Minute Newsroom 2026-08-21

Meta's assistant is now a Mac app that reads your screen and types into any window — and what it sees can train the model

Meta released its first desktop app for Meta AI on 19–20 August, a free beta for anyone with a Meta account. Two things make it more than a wrapper around the website. You can attach a specific window or screen from the prompt bar and have the assistant answer questions about what is on it, using Meta's Muse Spark model. And a system-wide dictation shortcut lets you hold a key, speak, and have the text land in whatever app you are in — mail, a document, a code editor. It also connects to Facebook and Instagram accounts to answer questions about ad campaign performance, and to Google Workspace for documents. The build is native rather than a packaged web page: AppKit and SwiftUI, 16MB installed, Apple silicon and macOS 15 or later. The caveat is in Meta's own terms. Material that passes through the assistant, including what is on screen during a sharing session and data pulled in from connected services, can be used to train Meta's models. Meta has also begun testing rate limits on heavier tasks, and the app lacks the deeper coding workflows and browser control that rivals offer.

Why it mattersThis is the third screen-watching desktop assistant to arrive in about a fortnight, after OpenAI's Mac agent and Cursor's push into hosting. The pattern is now clear: the assistants are moving off the web page and onto the operating system, where they can see everything rather than what you paste. That is genuinely more useful and it is a different privacy proposition, and the two are not separable — the same access that lets it answer a question about your spreadsheet is the access that makes the spreadsheet visible. Meta's version is the one to read the terms on most carefully, because its model training language is broader than the others'. Free, in this case, is a price.

✓ Verified · 4 sources

Read in the app — free, in 9 languages

Meta's assistant is now a Mac app that reads your screen and types into any window — and what it sees can train the model

Related stories