FR Newsletter

Automate

Muse

Meta's personal agent, which takes actions in the apps you connect.

Official siteVendor information

What the publisher documents

A personal agent published by Meta, announced on 8 September 2026. What separates it from an assistant is that it acts for you: Meta documents sending emails, booking travel, filling in forms, opening a browser and negotiating on your behalf, and it keeps working after you close the app. It is reached through the Muse app, on muse.ai, and directly inside WhatsApp.

The architecture is the part worth reading. Muse runs on a dedicated cloud virtual machine, the Muse Secure VM, which holds the agent, your data and the credentials of the services you connect. A second agent on the same machine, the Sentinel, is isolated from Muse at system level, and Meta's own sentence leaves no room: nothing Muse does reaches the internet unless the Sentinel approves it.

Which model answers you is not documented. Meta announced Muse Wonder on 12 October 2026 and wrote that it will soon power experiences in Muse, starting with subscribers. So the pages we read do not say what powers it today.

Your data

The training setting is the first thing to find, because Meta documents that it is on the first time you use Muse. You can change it at any time under Settings, then Data controls, and the change is unusual in one respect: it also applies to your previous interactions, where most privacy settings only govern the future. If you leave it on, Meta states that it still removes certain categories of personally identifiable information, such as names, email addresses, phone numbers and social security numbers, and that it separates the interactions from your account.

On advertising, the sentence is flat and worth quoting: Muse does not share your conversations or the data in your virtual machine with Meta's ad systems, and that holds even when your Accounts Centre includes other Meta products.

Credentials never reach the model. When you connect an app, usernames and passwords go into a secure credential store, and Meta documents three checks before one is used: that Muse is connecting to the intended service, that you connected or authorised that service, and that you authorised that type of action.

What Muse keeps is listed item by item: your interactions, the information it pulls from connectors such as emails and calendar events, the goals you set, the artefacts it creates, and the files on its system. Meta also states that it keeps backups of your agent's history. Deletion has two routes: ask Muse in the conversation to forget a subject or a set of interactions, or reset it under Settings, then Data controls, which Meta describes as permanently deleting everything, chat history, files and active tasks included. No retention period is published on the pages we read.

Plans and what you own

Muse is documented as free up to a usage limit, counted in Muse tokens per week, with paid plans that raise that weekly allowance; you subscribe with a Meta account, in the mobile app or on muse.ai. The documented conditions: be in a country where Muse is available, be 18 or the age of majority where you live, sign in with a Meta account, and accept Meta's subscription terms along with Muse's supplemental terms. Who owns what the agent produces is not addressed on the pages we read.

Documented limits

  • Probably not available where you are. Meta's own help centre states that the Muse app and website are not yet available everywhere, and that a feature available to you may not be available in your region. The four announcements we read name the US and Canada on 29 September 2026, then Singapore on 27 October 2026, and none of the four names France, Switzerland or the European Union; Meta publishes no country list.
  • Subscribing requires being 18, or the age of majority where you live.
  • Meta writes it plainly on its privacy page: no automated system is perfect, so review what Muse does and grant only the access you are comfortable with.
  • Websites, emails, files and connected services can carry instructions aimed at manipulating the agent, which Meta calls a prompt injection attack. Its answer is that the permissions and the important security checks work independently of the AI model, so they do not rely on Muse recognising a malicious instruction.
  • The Muse Confidential VM, the environment Meta describes as keeping your information private even from Meta, is announced as being designed, not as available.
  • The full safety report for the Muse Wonder models is announced as coming soon and is not published yet.

What we still have to verify

We have not tested Muse, and we could not: it is not available in the countries we work from. This page therefore reports documentation only, with no claim about quality, speed or style. Three things a test would have to settle: how long data actually stays, since no period is published; whether a reset really empties the virtual machine and the backups Meta says it keeps; and who owns what the agent produces, which the pages we read do not say.

Sources

  1. Meta Newsroom, Introducing Muse: The World's First Personal AI Agent Built for Everyone
  2. Meta Newsroom, The Future Is for Everyone: Muse for Small Business
  3. Meta Newsroom, Introducing Muse Wonder and Muse Wonder Full
  4. Meta Newsroom, Muse Is Now Available in Singapore
  5. Meta, Pages d'aide Muse, Comment Muse gère votre confidentialité, votre sécurité et votre protection
  6. Meta, Pages d'aide Muse, Comment gérer vos données Muse
  7. Meta, Pages d'aide, À propos des abonnements à Muse

Publisher pages read on November 1, 2026. We do not publish prices: they change often and the official page is the only reliable one.

Our test

We have not tested Muse yet. This page only reports what the vendor publishes. Our screenshots and verdict will appear here after the test.

Screenshot coming after our hands-on test
Muse : 1/3
Screenshot coming after our hands-on test
Muse : 2/3
Screenshot coming after our hands-on test
Muse : 3/3

Alternatives

Guides that use it