Meta's Muse Wants to Run Your Errands. The Real Question Is Whether You'll Let It
Meta has launched Muse, a personal AI agent that doesn't just chat, it actually does things. Send an email, book a trip, fill in a form, order the shopping. That's the pitch, anyway.
Muse went live in the US on September 8 through its own iOS and Android apps, on the web, and inside WhatsApp. Meta says it's coming to its smart glasses "soon", without putting a date on it. There's no word yet on a UK or European release.
What it can do
The big idea is that Muse plugs into the apps you already use: email, calendar, payments, health and fitness, smart home, shopping, dining, music and events. You give it a goal, it drafts a plan, and then it gets on with the work in the background, even when you're not in the app.
Some of the examples Meta has been showing off are genuinely handy, like turning a recipe reel into a grocery list or sending out party invitations. Purchases go through Link by Stripe, which adds buyer protection in case the agent orders something you didn't actually want. Meta also says Muse can negotiate on your behalf, which is either exciting or mildly terrifying, depending on how you feel about an AI haggling with a customer service bot.
According to Meta, Muse is modelled on OpenClaw, the open-source agent that's become something of a blueprint for this kind of tool.
The security setup
Meta clearly knows that trust is the whole game here, and it has put real thought into the architecture. Each Muse agent runs on its own virtual machine in Meta's cloud. A second AI, called Sentinel, keeps an eye on what Muse tries to send outside that machine and decides when it can access the internet. Passwords and payment details sit in secure storage, and the agent can use them without ever actually seeing them.
Permissions are granular, too. You could let Muse read your email but not send anything, for example, and you can revoke access at any time.
To its credit, Meta also admits that prompt injection, where hidden instructions buried in a web page or email trick an AI into doing something it shouldn't, is still an unsolved problem, and that Muse will sometimes make mistakes. It's refreshing to hear a company say that out loud at launch.
Pricing
TechCrunch reports two paid tiers at launch: Power at $20 a month and Maximum at $100 a month, with the pricier plans mainly giving you higher limits on delegated tasks.
Our take
On the tech side, Muse looks impressive, and the sandboxing approach is more careful than a lot of agent tools out there. But it's asking for your email, your money and your health data, and it's asking on behalf of Meta, a company whose track record on user trust is, let's say, complicated.
The technology may well be ready. Whether people are ready to hand Meta the keys to their digital life is a very different question, and one we'll get an answer to over the next few months.
