Meta Muse: what the "always-on" AI agent really does, and what it does with your data
Meta's Muse is not a bot that listens all day. It is an AI agent that keeps working in Meta's cloud after you close the app. Only in the US for now, and it trains on your chats by default.
Muse is Meta's new AI helper, and people call it "always on". That sounds like something that listens all the time. It is not. Muse is an AI agent that keeps working for you in Meta's cloud, even after you close the app. It can send emails, book things and shop for you. It is free to start, but only in the US, and by default Meta uses your chats with it to train its AI.
What is Muse?
Meta launched Muse on 8 September 2026. In its announcement, Meta calls Muse "a secure, private personal AI agent". An agent is an AI that does things for you, instead of only talking.
You talk to Muse like you talk to a friend: in the Muse app, on the website muse.ai, or in WhatsApp. You tell it what you need, and it gets to work. Meta gives examples like sending an email, booking a trip, filling in forms, selling a car or lowering a bill. Muse can also remember things you said once, and suggest ideas without being asked.
Muse runs on Meta's own AI model, called Muse Spark.
What "always on" really means
Here is the part that confuses people. Muse does not listen to you all day. "Always on" means that Muse keeps working when you are not looking.
Meta says it like this: for longer tasks, Muse "keeps working after people close the app". It comes back when something changes, or when it needs your OK, for example before it sends an email or buys something.
To do that, every user gets their own computer in Meta's cloud, which Meta calls the "Muse Secure VM". VM stands for virtual machine: a computer that only exists in software. Muse has its own web browser there, and it keeps your files and logins.
We found nothing on Meta's pages about a microphone that is always on. Meta says Muse is "coming soon" to its AI glasses, so that may change.
A guard called Sentinel
An agent that can send emails and buy things can also make mistakes, or be tricked by a nasty website. Meta's answer is a second AI, called Sentinel.
Sentinel runs on the same cloud computer, but separate from Muse. Muse may propose an action, but only Sentinel can let it reach the internet, Meta's security blog explains. For important steps, Sentinel asks you first, in a pop-up in the app.
Muse also never sees your real passwords. You hand them to a safe store, and Sentinel only adds them at the very last moment. Meta is honest that this is not perfect: tricking an AI with hidden instructions "remains an open problem", it writes.
What does Muse cost?
Muse is free to use, with a weekly limit. If you want more, Meta's help center lists two plans:
- Power: $20 a month, for 500 million "Muse tokens" a week.
- Maximum: $100 a month, for 3 billion Muse tokens a week.
Tokens are the small pieces of text an AI reads and writes. Meta does not say how much the free version gives you, and it says plans can differ per region and account.
Can you use Muse in the Netherlands?
Not yet. Meta's announcement says Muse is "rolling out in the US". Its help center says Muse is "not yet available in all locations".
We checked this ourselves from the Netherlands on 8 October 2026. The website muse.ai opened in Dutch and asked us to sign in. We did not make an account. In Google Play, the Muse app page showed over 1 million downloads, but no install button: only "add to wishlist".
Meta has not said when Muse comes to Europe.
Your chats are used for training
This is the most important thing to know. By default, Meta uses your conversations with Muse to train its AI. Meta's security blog says it removes personal details first, and calls this "a good default", because everyone then gets a better Muse.
If you do not want that, there is a simple switch in Muse's settings to turn it off. But you have to find it yourself: it is on unless you switch it off.
You can also tell Muse to "forget" things, and download what it remembers about you. Meta says that later this year, it will add a version where your data is locked with a key that only you have, "so not even Meta can access it". That is not there yet.
Muse and the ads you see
Meta promises that Muse does not share your chats or the data on your cloud computer with Meta's ad systems.
But there is a catch, and Meta explains it itself. When Muse visits a website for you, the website sees that as your visit. So if Muse buys a shirt for you, that shop might later show you an ad for it on Instagram. Bookings and Marketplace searches can also change the ads you see, Meta writes.
Should you trust it?
Meta has thought hard about safety. A separate guard, hidden passwords and approval steps are good ideas, and Meta explains them in more detail than most companies.
But a few things are worth remembering:
- Your "own" computer still runs in Meta's cloud.
- Training on your chats is on unless you switch it off.
- Websites Muse visits can still show you ads.
- The key that only you hold is promised, not here.
Our advice: if Muse comes to Europe, give it small, low-risk jobs first, check what it asks permission for, and switch off training if you do not want to share your chats.
How we checked
On 8 October 2026 we read Meta's announcement, its security blog and its help page on subscriptions. The screenshots are our own captures of those pages, cropped only. For the prices we made our own chart, because a cookie window covered the help page.
We could not test Muse itself, because it is not available in the Netherlands, and we do not create accounts for our tests. We only looked at muse.ai and the Google Play page from the Netherlands. So everything about how Muse works comes from Meta's own words, not from our own use.
The charts were made by us. None of the images was made by AI.