Meta announced Muse on 8 September 2026, and the pitch is not that it will answer your questions. It is that it will do things for you: read your email, book your travel, fill in your forms, negotiate on your behalf, and pay for the results with your saved card.

Muse is a personal AI agent, rolling out across the United States on iOS, Android, the web and WhatsApp, with support for Meta's AI glasses promised later. It is the most direct consumer expression yet of Mark Zuckerberg's stated vision that everyone will have an AI assistant working for them.

An agent, not a chatbot

The difference Meta is selling is autonomy. Muse keeps working after you close the app, makes suggestions nobody asked for, and remembers your preferences across sessions. It will turn a saved recipe reel into a shopping list, hunt for a lower bill, complete a form, send invitations and make purchases — acting rather than replying.

Under the hood it runs on Muse Spark, which Meta describes as its "most capable model to date, built for real-world agentic work."

There are three tiers: a free tier with metered usage, a $20-a-month "Power" plan and a $100-a-month "Maximum" plan.

The part Meta wants you to look at

An agent that holds your inbox and your card raises an obvious question, and Meta has built an architecture aimed squarely at it. Each user gets an isolated virtual machine. A permissions layer called Sentinel sits between Muse and the internet, distinguishing read access from write access and issuing time-limited authorisation for actions. Meta says credentials are stored so that the agent cannot see raw passwords or card numbers, and that Muse conversations are kept separate from its advertising systems. A "confidential" mode, in which Meta itself cannot observe what happens in your workspace, is promised for later.

The part it would rather you didn't

That architecture collides with the company presenting it. Meta's privacy record includes FTC settlements in 2011, 2019 and 2023, the Cambridge Analytica scandal, and a multistate child-safety settlement of up to $18 billion approved in August 2026. And the isolation is real, but some of the strongest reassurances are promises rather than architecture: keeping Muse conversations out of the ad systems is a policy commitment Meta could revise, and the "confidential" mode in which Meta cannot observe your workspace has not shipped yet. Those are statements about what Meta will do with your data, not walls that stop it.

There is a second risk that has nothing to do with trust in Meta, and it is inherent to the whole category: an agent that can fill a form and complete a purchase can also fill the wrong form and complete the wrong purchase, at speed. Handing over the ability to act is also handing over the ability to act wrongly, faster than you can intervene.

Why it matters

Muse is the clearest consumer bet so far that people will give an AI not just their questions but their calendar, their inbox and their card. The technology to book a table or clear a bill autonomously is largely here. Whether it gets adopted is now less a question of what the agent can do than of whether people trust who is holding the keys — and Meta is asking for that trust from a standing start it dug itself.