Meta launches Muse - a personal AI agent that does the tasks itself

On 8 September Meta launched Muse - a personal AI agent that differs from an ordinary chatbot in that it actually does the work. You give it a goal or a task, it makes a plan, gets going, and carries on after you've closed the app. It comes back when something changes, or when it needs your approval.

Muse runs on its own virtual machine in Meta's cloud, with its own browser that you can watch. Passwords and payment cards sit in secure storage that the agent can use but not see, and a separate guard agent has to approve everything Muse sends out onto the internet. You get a full log of what it has done and intends to do.

Getting started

1. Check whether you can - so far Muse is only rolled out in the US, and only to users aged 18 and over. It isn't available in Denmark yet.
2. Pick your entry point - there's an app for iOS and Android, a web version at muse.ai, and you can message the agent directly in WhatsApp. Support for Meta's AI glasses is on the way.
3. Set up your agent - you give it a name, can choose an avatar, and decide how it should communicate with you.
4. Connect what it may see - email, calendar, Instagram, Facebook, health data and finances are connected one at a time, and you decide the level: may it only read your email, or also send on your behalf? You can disconnect at any time.
5. Choose a plan - most things can be done for free. If you want to use it more, there are two paid tiers at 20 and 100 dollars a month, and a payment card has to be registered at sign-up.

What can you use it for?

Muse is meant for the practical clutter of everyday life rather than for writing texts:

• Calendar and email - move appointments that clash, book a table, set reminders and chase people who haven't replied.
• Longer projects from start to finish - research, comparisons, the tedious back-and-forth and the actual ordering. Meta highlights examples like selling the car for a better price, negotiating a bill down, or planning and booking an entire trip.
• Shopping - compare prices, watch for price drops and complete the purchase once you've approved it. Payment goes through Link from Stripe with a one-off card, so your real card details aren't passed on to the shop.
• Finances - keep track of expenses, review subscriptions and cancel what you don't use. It asks before it cancels or spends money.
• Habits and training - a running programme or meal plan that's adjusted as the rest of the calendar shifts.
• Small practical things - fill in forms, such as a permission slip for a school trip, or buy cinema tickets.

It also remembers what you've told it. A recipe you saved on Instagram can become a shopping list, and it can remember your friends' allergies before sending out the invitations. You can always ask it to forget something specific again.

Worth considering

That kind of help costs access. Muse asks for far more personal information than an ordinary chatbot - email, calendar, payment and health - and several observers point out that this will be the real test: whether people trust it enough to give an agent that access. Meta itself stresses that conversations and data from your machine aren't shared with the advertising systems, that you can opt out of your conversations being used to train the models, and that later this year a version is coming where the whole machine is encrypted with a key only you hold.

Read Meta's own announcement at about.fb.com.