Meta AI Lands on the Mac With Screen Sharing, and a Permissions Prompt Worth Reading

Meta shipped its first Mac app on Wednesday, and the most consequential thing in it is not the chat window. It is the macOS dialog box that appears the first time you ask the assistant to look at what you are working on.

The Meta AI desktop app for macOS went live on August 19, 2026, as a version 1.0 beta. It is a 16MB download that runs natively on Apple silicon Macs running macOS 15 or later, built on an AppKit and SwiftUI shell with WebKit for richer chat content rather than shipped as an Electron wrapper or a repackaged iPad binary. It is available in select regions and requires a Meta account, which in practice means a Facebook, Instagram, Messenger or WhatsApp login.

Three desktop features distinguish it from the Meta AI website and mobile app. The first is window attachment: point the assistant at a single open window and ask questions about it. The second is Quick Invoke, a compact composer summoned with Option-Space that floats over whatever application is in front, with the option to hide the app from the Dock entirely so the shortcut becomes its primary access point. The third is system-wide dictation, which lets a user hold a shortcut, speak, and have the transcribed text land in a mail client, a document, or a code editor.

Window attachment is the feature that requires reading the fine print. To read a window, the app needs two macOS permissions: Screen Recording and Accessibility. With both granted, Meta AI reads the visible text in the selected window and captures a screenshot of it, then answers questions, suggests edits, or drafts content based on what it sees. The design gathers context rather than taking action. Meta AI can interpret a shared window, but it cannot click buttons, edit files, or drive applications on a user’s behalf.

The business pitch, and the data behind it

This is not a general consumer play. Meta is aiming the app at small business owners and content creators who run Facebook and Instagram presences, and the launch arrived alongside business integrations spanning Mac, mobile and web. People with professional Facebook or Instagram accounts can connect Meta AI to their account analytics, their Meta advertising campaigns, and Google Workspace. The assistant can read post reach, likes, shares and saves and recommend what to publish next, schedule recurring tasks such as weekly performance reports, and generate decks, docs and spreadsheets.

Meta’s own framing is blunt about the moat it believes it has. “You can ask Meta AI questions about your business and get answers drawn from the context only Meta has, like your account engagement and ad performance,” the company wrote in its announcement post. “Beyond your own accounts, you can get publicly available insights about comparable brands and how they are building their presence on Meta.” The pitch for the desktop app itself is similarly direct: users can “point Meta AI at any window on your desktop to work through your tasks or use your voice to dictate in any app.”

The app is free. Meta has been steadily introducing rate limits on more compute-intensive features, and higher ceilings come with a Meta One subscription.

The permission is the product

Strip away the branding and this is a land grab, and a late one. Every major lab now ships a desktop client. Google’s Gemini already supports window sharing on macOS. ChatGPT and Claude go further, taking control of the machine rather than merely reading it. Meta arrives with a version 1.0 that reads what you hand it and stops there. Reviewers who tested the build called browser use and computer control conspicuous gaps, and noted that Meta AI still lacks the deep coding workflows its rivals ship, even though Meta released its own terminal coding agent, Muse Code, earlier this month.

So the differentiator is not raw capability. It is the data Meta already holds about a creator’s audience, and that is where the story gets uncomfortable. Meta’s generative AI privacy documentation states that the company trains on “Information shared on Meta Products including things like posts or photos and their captions, and your interactions with features that are part of AI at Meta.” It also draws one explicit line: “We do not use the content of your private messages with friends and family to train our AIs unless you or someone in the chat chooses to share those messages with our AIs.”

A window you attach to a Meta AI conversation is, on any plain reading, an interaction with a feature that is part of AI at Meta. Meta has not published a desktop-specific carve-out saying otherwise, which leaves concrete questions unanswered. How long are captured screenshots and extracted window text retained. Are they stored on Meta servers at all, or discarded after the turn. Are they excluded from model training. Are Google Workspace documents pulled through the connector treated differently from chat content. Do business administrators get any control, audit trail, or revocation path beyond an individual user’s own settings. None of that is spelled out in the launch materials.

A second technical point is worth flagging. Accessibility is one of the broadest grants on macOS. It exists for assistive technology, and it lets an app observe and interact with interface elements across the entire system. Meta says the app only reads the window you attach. The permission itself does not enforce that scope. Both permissions live in System Settings under Privacy and Security and can be revoked at any time, which is the practical advice for anyone testing the app.

To be fair to Meta, the trade-off is not unique to it. Any screen-aware assistant asks for the same access. The difference is business model. Meta monetizes attention and targeting, and it is the only one of the four major desktop assistants asking creators to plug their advertising performance data into the same tool that can see their screen.

Three things to watch. Whether Meta publishes desktop-specific retention and training commitments for captured windows, which would defuse most of the criticism at low cost. Whether a Windows client and computer-control capabilities follow, since the current app is a context reader in a market moving rapidly toward agents. And whether European regulators take an interest, given that questions about consent, profiling and cross-platform data use have already trailed Meta’s personalized AI agents under EU privacy rules. The app is small, fast and free. The permissions it asks for are neither small nor easily unwound.