Meta launches Muse, an AI agent that books, fills forms and buys

Meta announced Muse on 8 September 2026, describing it as a secure, private personal AI agent that proactively helps with people's goals and suggests ideas. The Muse personal AI agent is built to execute rather than converse: Meta says it can send an email, book travel, open a browser, fill out forms, negotiate on a person's behalf and complete a checkout, and that it keeps working after the app is closed.

It is rolling out in the United States only, on iOS, Android and muse.ai, with AI glasses named as coming soon. Meta says it is free for most of what people need, with subscription plans for people who want to do more. The announcement contains no business tooling, no advertiser product and no brand controls of any kind, which is the fact that matters most to anyone selling online.

What Meta announced on 8 September 2026

Meta's newsroom post introduces Muse as an agent that runs on its own infrastructure, powered by a model Meta calls Muse Spark and described as its most capable model to date, built for real-world agentic work. Meta also says Muse was designed so that talking to it works like messaging another person, either in the Muse app or directly in WhatsApp.

The framing throughout the post is about doing rather than answering. Meta writes that people tell Muse what needs to get done and it takes action. For a bigger goal, Meta says Muse helps develop a personalised plan, coordinates time and resources, and then advances the work on its own.

What Muse actually does, in Meta's own description

Meta lists a set of concrete capabilities rather than a general promise. Muse can handle discrete tasks such as sending an email or booking travel. It can open a browser, fill out forms and negotiate on a person's behalf. Meta's examples of the payoff are selling a car for more, lowering a bill, and adjusting a training plan as the rest of someone's life shifts.

Memory is part of the design. Meta says Muse remembers what matters to a person, can make suggestions unprompted, and can act on details the person mentioned only once. The worked example given is turning a recipe reel saved on Instagram into a grocery list, suggesting a menu for a dinner party, and remembering friends' dietary restrictions before sending invites. Meta also says a person can tell Muse to forget specific things it has learned.

It keeps working after the app is closed

For tasks that take longer, Meta says Muse keeps working after people close the app and comes back when something changes or when it needs approval, giving sending an email or making a purchase as the moments where it asks first. That asynchronous behaviour is what separates this from a chat session, and it is also what makes the agent a repeat visitor to websites without a person watching the screen.

The security architecture Meta named

Meta gave the security design more space than the features, and named the components rather than describing them vaguely. Muse runs inside Muse Secure VM, a dedicated virtual machine in the cloud that houses both the agent and the person's data, contained so that no one else's agent can reach it. Credentials for any service a person connects are stored there.

A separate Sentinel agent runs on the same machine, kept apart from Muse at the system level. Meta states that nothing Muse does reaches the internet unless the Sentinel approves it, and that it asks the person for permission when needed. Meta also says Muse has no visibility into passwords or payment methods, including passwords a person types into the browser themselves, because credentials go into secure storage that Muse can use without seeing.

On top of that, Meta describes an audit trail showing everything Muse has done and plans to do, per-app access controls where the person chooses which apps Muse connects to and how much access it gets, the ability to change access or disconnect a service at any time, and an opt-out from having interactions used to train Meta's AI models. Meta states directly that Muse does not share a person's conversations or the data in their VM with Meta's ad systems. Later this year, Meta says it will introduce Muse Confidential VM, where the whole VM including data and conversations is encrypted with a key only the person holds, so not even Meta can access it.

How Muse pays for things

Checkout runs through Link, built by Stripe. Meta says Muse is the first AI agent covered by Link's purchase protections, listing free coverage for damaged or lost items, price drops, no-fee returns and a return guarantee on eligible purchases. Link's wallet for agents generates a one-time-use card so the person's real card details stay hidden. Meta says Shop Pay is coming soon as another way to pay, along with 1Password support so Muse can use logins a person already has.

The announcement names its parts, which is worth setting out plainly.

What the announcement did not say

The omissions matter more than usual here, because a launch this broad invites people to fill the gaps with guesses.

• No country other than the United States is named, and no timeline is given for any other market. Thailand is not mentioned.
• There is no advertiser, brand or merchant product. Meta names no way for a business to appear to Muse, to be preferred by it, to pay for placement in it, or to opt out of it.
• Meta does not say how Muse identifies itself when it opens a browser and visits a website, nor whether site owners can detect or control agent visits.
• No pricing figures are given for the subscription plans, only that they exist alongside a free tier.
• No date is given for the AI glasses version or for Shop Pay and 1Password support beyond coming soon, and Muse Confidential VM is placed only later this year.
• Nothing is said about how Muse chooses which merchant, airline or hotel to transact with, or what it uses to compare options.
• No usage numbers, no waitlist detail and no measurement of how well the agent completes tasks are included.

Why an agent that buys is a marketing problem, not a gadget story

Strip out the product language and one fact remains: Meta has shipped software that opens a browser, fills in a form and pays, in a product Meta itself walls off from its own ad systems. Read the two together. A buyer using Muse does not see a display impression, does not scroll a feed, and does not respond to a retargeting sequence. The statement that Muse does not share conversations or VM data with Meta's ad systems means, on its face, that there is no advertising route into this behaviour at launch.

What is left is the data. An agent comparing options works from what a site publishes: product attributes, prices, availability, shipping and returns terms, and whatever else it can parse. Meta did not describe how Muse evaluates a page, so nobody should claim to know the ranking mechanics. The defensible conclusion is narrower and still useful. If the buyer is software, the parts of a page that were written for a human skimming a hero image do less work than the parts a machine can read cleanly.

This is the same discipline that generative engine optimisation already asks for, which is why it is worth treating as one body of work rather than two. Making a page quotable by an answer engine and making a page legible to a purchasing agent both come down to publishing facts in a structured, unambiguous, crawlable form rather than burying them in scripts and images.

What a marketer can actually do about it now

Very little of the useful work is Muse-specific, which is the point. The checks worth running are the ones that pay off across every machine reader.

• Confirm that price, stock status, shipping cost and delivery time render in the page HTML rather than only inside a client-side widget.
• Put returns, warranty and cancellation terms in plain text on a stable URL, not in a PDF or an accordion built entirely in script.
• Check that product structured data matches what a customer sees, since a mismatch between feed and page is the kind of thing an agent surfaces as an error rather than forgiving.
• Look at how many steps a purchase takes and how many of them require a human gesture that an agent cannot perform.
• Decide, as a policy question, what your position is on agent-completed purchases and agent-initiated contact through your forms, because Meta says Muse fills forms.

For businesses running ecommerce marketing, the practical exposure is the checkout and the product data, in that order. Meta describes Muse checking out with a one-time-use card through Link, which means the transaction can look ordinary to the merchant while the decision that led to it happened somewhere no ad reached.

What this means for Thai marketers

Muse is not available in Thailand. Meta named the United States as the launch market and gave no other country and no timeline, so any Thai rollout is unannounced rather than imminent or ruled out. Nothing in a Thai media plan needs to change this month because of Muse.

The reason to pay attention anyway is that Meta is a platform Thai advertisers already depend on heavily, and this is the clearest signal yet of where the company thinks consumer behaviour is heading. A product built on the premise that people will delegate booking and buying, shipped with an explicit separation from the ad business, is worth reading as a direction rather than a feature.

The preparation that makes sense is the kind that has value whether or not Muse ever launches in Thailand. Clean product data, crawlable prices, plain-text policies and a checkout that does not depend on unusual interactions all improve conversion for humans today and readability for machines later. That is a rare case where the hedge costs nothing.

Frequently asked questions

Is Muse available in Thailand?

No. Meta said on 8 September 2026 that Muse is rolling out in the United States on iOS, Android and muse.ai, with AI glasses coming soon. The announcement names no other country and gives no timeline for any other market, so there is no stated plan for Thailand either way.

Can brands advertise inside Muse or pay to be recommended?

Not according to the announcement, which describes no advertiser or business tooling at all. Meta states that Muse does not share a person's conversations or the data in their VM with Meta's ad systems, and names no placement, partnership or merchant programme for businesses.

Will Muse buy things without the user knowing?

Meta says no. The announcement states that Muse checks with the person before sensitive actions such as sending an email or making a purchase, that a separate Sentinel agent must approve anything Muse sends to the internet, and that Muse shows a complete audit trail of everything it has done and plans to do.

Does Muse have access to my card details?

Meta says Muse has no visibility into passwords or payment methods. Credentials a person shares go into secure storage that Muse can use without seeing them, and checkout runs through Link, built by Stripe, whose wallet for agents generates a one-time-use card so real card details stay hidden.

Do I need to change my website for AI agents now?

Nothing about Muse forces a change today, since it is a United States launch with no advertiser or publisher programme attached. The work that holds up regardless is publishing prices, stock, shipping and returns terms as readable text rather than script-only widgets, which helps human buyers and machine readers alike.

What to do with this

Read the announcement yourself before acting on anyone's summary of it, including this one, because the operative details are Meta's own words about scope and limits. It is published on Meta Newsroom.

If you sell online and want to know how your product data and checkout would look to a machine buyer rather than a human one, that is a concrete audit with a concrete output, and it is worth doing while the answer still only affects conversion rather than revenue.