Meta Muse App: How to Set Up and Use Muse, Meta’s Personal Agent

Muse is not another chatbot. It is an agent: you give it a task, and it goes and does the task, sometimes after you have closed the app. Meta launched it on 8 September 2026 and describes it sending emails, booking travel, filling out forms and making purchases. That is useful, and it is also exactly why setting it up carefully matters more than it does for a tool that only talks back.

This guide covers who can get the Meta Muse app, what it asks for, how to connect your accounts without handing over more than you need, and three first tasks worth trying. It sticks to what Meta and the press have actually reported, so it describes steps in general terms rather than inventing screens.

Muse, Muse Spark and Meta AI are different things

The names overlap, which causes most of the confusion. Muse Spark is the model, the underlying AI that also runs the Meta AI app and meta.ai. Muse is the new agent app built on top of it. If you want the background on the model itself, what Muse Spark is covers it separately.

So if you have used Meta AI to ask questions, you have used the same family of model. What changes with Muse is that it can act: open a browser, fill in a form, and send something on your behalf once you say yes.

Who can get it

As of September 2026, Muse is rolling out in the US on iOS, Android and muse.ai, with support for Meta’s AI glasses described as coming soon. Outside the US, there is no launch date to report.

• —On iPhone, the App Store listing is called Muse from Meta, which is the name to search for.
• —On Android and on the web at muse.ai, it is the same agent. "Rolling out" means it may not appear for every US account on day one.
• —You can also talk to it through WhatsApp as well as in the app, according to Meta’s announcement.

What it asks for before you start

There are three plans: Free with a usage meter, Power at $20 a month and Maximum at $100 a month, according to TechCrunch. The same report says a payment card is required to start, even on the free plan. Meta expects most people to stay on free.

Go in expecting that card step, and decide beforehand whether you are comfortable with it. The differences between the three plans are set out in Meta Muse pricing, so this guide will not repeat them.

The free plan has weekly usage limits, so a handful of small tasks is a better first week than one enormous one. The same report says you can change the agent’s name and avatar, which is cosmetic and safe to skip.

Connect one app at a time, with the least access

Muse works by connecting to your other services, and those connections are opt-in one by one: email, calendars, payments, health and fitness, smart home, dining, shopping, music and events. Where a service has no connection, it uses a browser instead.

The important detail is that you choose how much access each connection gets. Meta gives the example of email access that can read only, versus access that can also send. Start with the narrower option every time.

1. 01Connect one service only, for the task you actually want to try this week. Calendar and read-only email are sensible first choices because nothing leaves your account.
2. 02Pick the lowest level of access that lets the task work. You can widen it later; you cannot take back an email already sent.
3. 03Leave payments unconnected until you have watched it handle a few tasks without money involved.
4. 04Look at the privacy choices early. Meta says you can opt out of your interactions being used to train its models, ask Muse to forget things, and disconnect a service at any time.

When you do connect payments, Meta says they go through Link by Stripe, and that Muse has no visibility into your passwords. Shop Pay and 1Password support are listed as coming soon, not available yet. The fuller picture on data and permissions is in is Meta Muse safe.

The approval step is where your judgement goes

Muse asks for your approval before it sends an email or makes a purchase. Meta also says the model version it runs, Muse Spark 1.3, asks clarifying questions when a request is ambiguous and confirms before consequential actions.

That approval is only as good as the attention you give it. An agent that has done ten tasks correctly makes it tempting to tap yes on the eleventh without reading. Treat each approval as the moment you are signing your name to what it did.

The approval prompt does not check the work. You do. Read what it is about to send as if a stranger had written it.

Three first tasks worth trying

Good first tasks are small, easy to check, and cheap to get wrong. If you are not sure which of your own chores qualifies, finding the repetitive part is a quick way to pick one.

1. 01Turn saved recipes into a shopping list. Meta’s own example is turning saved Instagram recipe reels into a grocery list. Nothing is sent or bought, and you can check the list against the recipes in a minute.
2. 02Plan something with your calendar. Ask it to find a free evening next week and draft a plan around it. With read-only calendar access, the worst outcome is a bad suggestion you ignore.
3. 03Draft and send one invitation. Party invitations are another example Meta gives. This is the first task that sends something, so it is the one where you practise reading the approval properly.

A worked example: briefing it well

Most disappointing agent results come from a vague brief, not a weak model. The habits in writing a prompt that works first try apply here, with one addition: tell it what it must not do.

Weak: "Invite my friends to a birthday dinner on Saturday."

Better: "Draft an invitation email for my birthday dinner. Saturday 26 September, 7:30pm, at our flat. Send it to these six people [names and addresses]. Ask them to reply by Wednesday and mention any dietary needs. Keep it short and friendly, no more than five lines. Do not send it until I have approved the draft, and do not book or buy anything."

The better version fixes every detail the agent would otherwise guess: the date, the time, the guest list, the deadline and the tone. It also draws a boundary around spending, which matters for a tool that can make purchases. For more briefs like this, tried against the model directly, see the Muse Spark tutorial.

What to check before you approve

• —Recipients. Every name and address, one by one. A wrong autocomplete is the easiest mistake to miss and the hardest to undo.
• —Dates, times and amounts. These read just as confidently when they are wrong, and AI models are well documented to state errors fluently.
• —Anything it added that you did not ask for, such as an extra line, a promise, or an attachment.
• —For purchases, the total, the seller, and whether the item matches what you meant rather than what it found.
• —Whether this is something you would be happy to have sent under your name without explanation.

When the draft contains facts you cannot easily confirm, the habits in checking an AI answer when you are not the expert still apply to an agent.

What it is not yet good for

Some limits are simply about launch timing. It is US only, the glasses support is not out, and the Muse Confidential VM with keys held by the user is promised later rather than available now.

Others are about judgement. Meta says Muse can negotiate on your behalf and lower bills, but a negotiation involves trade-offs only you know about, and a wrong concession is not an edited draft. Hold those back until you have seen how it handles simpler tasks. The same goes for anything where a mistake costs real money or a relationship, for the reasons set out in what AI is actually bad at.

Work accounts deserve their own caution. Connecting an employer’s email or calendar to a personal agent is a decision for your employer’s policy, not for the app. The split between what an assistant can take on and what stays with a person is covered in using AI as an executive assistant.

It is also fair to weigh who is running it. Meta says Muse runs on a dedicated cloud machine, with a separate agent checking actions that go out to the internet, and that it does not share your conversations or the data on that machine with its ad systems. TechCrunch also points out Meta’s privacy record, including a $5 billion FTC penalty in 2019. Neither fact settles the question alone. Both belong in your decision about how much to connect.

What to do this week

1. 01If you are in the US, install Muse from Meta or open muse.ai, and decide in advance how you feel about the card step.
2. 02Connect one service with the narrowest access that works.
3. 03Run the shopping-list task, then the calendar task, then one invitation with a full brief.
4. 04Read every approval as if you were about to sign it, and only widen access once it has earned it.

The skill that makes an agent useful is not the app. It is briefing precisely and checking the result before it goes anywhere. That is what Coursium teaches, on iPhone, so you can stay ahead of AI whichever tool you end up using.