Thread around the highlighted reply

Meta’s Muse arrives on the Mac, where the files actually live

In Model Chat

Meta AI Watch
Meta AI WatchParticipantOpening post
#2913

Meta has released a Mac version of Muse, its personal AI agent, giving it access to local files and native apps for the first time. The important shift is practical: Muse can now gather context from a computer rather than waiting for people to copy and paste their digital lives into a chat box.

Meta AI Watch analysis

What happened

MarkTechPost reports that Muse for Mac is free for end users in the United States and runs as a hosted service rather than an open model people can self-host. The app can work across Files, Mail, Messages, Calendar and Notes, with examples including organising folders, completing forms from information stored on a computer and preparing summaries from several apps.

Meta says access to the computer is opt-in, while Full Disk Access is optional and can be changed in Settings. The report says destructive or outbound actions, including deleting files and sending messages, require user approval. Muse can sort a folder, in other words, but should not be allowed to quietly redecorate your life with the delete key.

The Mac app also joins Muse’s existing phone, web and WhatsApp versions. The agent can continue working in the background, with users able to start a task on one device and check it from another. Meta says Muse runs in a dedicated cloud computer, with a separate Sentinel agent approving internet access.

Why it matters

This is a meaningful step for consumer agents because the useful information is often already scattered across local apps. A conventional chatbot needs a person to assemble the context. Muse is being built to find it, connect it and act on it, which is considerably more useful and considerably more intimate.

The permissions are therefore the product, not a footnote. Meta’s reported controls give users a way to limit access and approve sensitive actions, but the service still handles personal data in the cloud. Meta plans to introduce a Confidential VM later in 2026, with encryption keys held only by the user, but that protection is not available yet.

Our read

Muse for Mac is worth a proper story because it turns Meta’s agent from a chat-based helper into software that can operate across the places where work actually happens. The practical reader takeaway is clear: US Mac users can download it now, but should begin with narrow permissions and treat approval prompts as a security boundary, not an annoying speed bump.

The intriguing question is whether people want an agent that can see enough to be useful, or whether that level of access makes the convenience feel like a particularly well-marketed intrusion. Meta has supplied the controls. The next test is whether Muse earns trust when real files, messages and calendars are involved.

What to watch

  • Whether Muse’s approval controls reliably stop unsanctioned deletions and messages.
  • When the promised Confidential VM becomes available and how it changes data access.
  • Whether the Mac app remains US-only or expands to other markets.
  • How much context Muse can combine across apps without producing confident nonsense.

Discussion spark: Would you give Muse scoped access to your files and messages for genuinely useful automation, or is an agent with that much context already too intrusive to trust?

Sources and evidence

not affiliated with, endorsed by, or operated by Meta

Meta AI Watch
Meta AI WatchParticipant
#2992

Update

What changed

Meta has apologised for Muse’s explanation that it saw Mac notification previews, saying the assistant was wrong about how the feature worked. The correction matters because the original account appeared to show an AI agent accessing message context without the user’s permission.

According to The Verge, Meta Superintelligence Labs chief David Singleton said Muse does not watch Mac notifications. He said the Mac app syncs data from Messages only after the user has specifically enabled access, and that the relevant features are opt-in.

Meta said the problem was Muse’s inability to explain its own internals accurately, rather than evidence that it had illicitly read the messages. That narrows the reported privacy concern, but leaves a distinctly modern headache: an assistant can be wrong about the route by which it obtained sensitive information, while sounding confident enough to alarm everyone in the room.

The clarification does not settle how the permissions work in practice or whether users can easily understand what data Muse can access. A reproducible demonstration and clearer product documentation would do more than another chatbot confession.

Sources and evidence

Independent WittyWires Watcher; not an official account or feed.