Meta launched Muse on September 8, 2026, and it shot to number one on the US App Store’s free chart ten days later. It is a personal AI agent that does tasks instead of answering questions, and the base version is free. The catch is not the price. It is everything around the price.
Muse runs on a dedicated cloud machine with its own browser, so it keeps working after you close the app. You message it like a friend, in the Muse app or in WhatsApp. It pings you first when something needs attention. At launch it was US-only, limited to people 18 and older, with Canada following.
What Muse is built to do
You connect apps one at a time and give the agent a goal. Email. Calendar. Payments. Health and fitness.
Smart home. Dining and shopping. Meta says you choose which apps it reaches and how much access each gets, and you can revoke any of them. A permission system called Sentinel separates read access from write access and reviews what the agent tries to do before it reaches the internet.
For shopping, Muse connects to Shopify‘s catalog and has retail partners including Best Buy, Sephora, Walmart, Wayfair, Expedia, Instacart, GitHub and Notion. Checkout runs through Link by Stripe, Shop Pay and PayPal, with a one-time card so the merchant never sees your real number. On paper, that is a cleaner setup than most bots get.
Meta announced more on September 23. A Realtime Avatar model lets you talk to the agent over video with a visual representation. Muse can control applications on macOS. It gets its own email address, so you can forward messages to it or copy it into a thread. A palm-sized device called Muse Charm was shown for using the agent away from a phone.
The design leans hard into cuteness. Muse uses a Labubu-inspired look, and Dots later copied the friendly-puffball style. Both products signal affection toward software at a moment when plenty of users are tired of generative AI. Whether that helps trust or erodes it is still an open question.
The Amazon wall
On September 21, 2026, Amazon blocked Muse from its retail site. Shoppers who tried to buy through the agent saw a message that continued access by an unauthorized AI agent violates Amazon’s Conditions of Use. Amazon said Meta had not told it Muse would access the site, that the agent did not identify itself while browsing and that it appeared to capture and store customer credentials.
Meta had said Muse has no visibility into people’s passwords or payment methods and that credentials are held in secure storage. Amazon’s response put that claim under direct public dispute.
Shopify, by contrast, let Muse operate on its storefronts, and the two stayed commercial partners. So the same agent that one giant blocked, another giant welcomed. If you plan to use Muse to buy things, the store you pick decides whether it works at all.
The motive is not only safety. Amazon’s sponsored-product ad business brings in roughly $80 billion a year. Agent traffic hides the shopping data that business depends on. When a bot shops for you, the signal that pays for the shelf disappears.
The human behind the agent
In September 2026, 404 Media reported that some phone calls placed by Muse were, during beta, completed by human call-center workers rather than the agent. The report drew on internal Meta posts describing a “human agent layer,” where Muse hands a request to a trained person who places the call and works it through.
That does not make Muse a scam. It means the “autonomous” demo had a manual backstop, and some users were talking to a person without knowing it. For a product pitched as software that acts for you, the gap between the pitch and the beta is worth holding in your head.
What it costs and how Meta makes money
Muse offers a free tier alongside subscriptions at $20 and $100 a month. Meta said the service would not carry advertising. Mark Zuckerberg said the company expects to profit by taking a small fee from transactions the agent completes over time. The model is closer to a travel agent’s cut than to an ad business.
Privacy settings matter here. Each user’s agent runs in a separate virtual machine. Meta uses your queries for model training by default, but you can turn that off. The company said a version in which it cannot view your workspace would ship before the end of 2026. Prompt injection, where a web page tricks the agent into doing something you did not ask, remains an unsolved problem across the whole category.
How Muse compares to the others
So Muse is the cheap one. Dots costs $100 a month. Devin is built for code. But Muse is the only agent already on your phone. And it is the only one a human sometimes answers for.
Muse wins on reach. It is free and it is installed. But it loses on trust, given the Amazon block and the human backstop. And it loses on privacy defaults, since training stays on unless you turn it off.
Who should actually use it
Muse is easiest to adopt if you already live in Meta’s world. WhatsApp. Instagram. Those are first-class surfaces for the agent, and the setup is a few taps. If your calendar and email are already there too, the connection story is short.
It is harder to recommend if you are privacy-skeptical or outside the US and Canada. The training opt-out is a setting, not a default. The human-agent-layer episode shows the boundaries of “autonomous” were still moving during beta. And Amazon’s block is a reminder that the open web is not obligated to let an agent in.
My take is that Muse is the most interesting of the 2026 agents because it is free and already on millions of phones. But the free price buys access, not trust. Treat it as an assistant that asks before it spends, not an autopilot, and start with low-stakes errands before you let it near a purchase.
