Ask a chatbot for running shoes and it can explain cushioning, compare brands and suggest a shortlist. Ask Muse and it can open the shops, filter for your size, keep looking while you do something else and stop at checkout for your approval.
That change — from giving an answer to carrying out the steps — is the point of Muse. Meta launched it in the United States on 8 September 2026 as a separate app, a web service and a contact inside WhatsApp. Within ten days it had reached the top of the US free iPhone chart.
What Muse can do
Muse has its own computer in Meta's cloud. It includes a browser, files, storage and a terminal. The agent can search the web, move between sites, fill forms, write code and build a tool when a task needs one. Connectors let it work with services such as email, calendars, Instagram and Facebook.
It can also keep going after you close the app. A task may finish in the background, wait for a change or return when it needs a decision. Longer goals live in a Goals view, and Muse can suggest a next step without waiting for another prompt. Its activity log shows what it has done, what it is doing and which permissions it holds.
Sensitive actions are meant to stop at a clear approval screen. Muse asks before it sends an email or completes a purchase. Passwords and payment details sit in protected storage that the main agent cannot read. A separate system called Sentinel decides whether outside actions and network requests are allowed.
It is not the only AI agent
Muse is different from the old picture of ChatGPT or Claude as a box that only writes replies. It is much less different from their current products.
On eligible plans, ChatGPT Work can use a separate browser in the cloud, sign in to websites, click, type, complete multi-step tasks and continue after the user leaves. Claude can work through connected apps, a built-in browser or the user's Chrome, create finished files and, in beta on Pro and Max desktop plans, operate local apps. Both can pause before a consequential action.
Muse's distinction is the whole bundle. Meta presents it as one continuing personal relationship, complete with a name and avatar, memory, goals and messages it starts itself. Every user receives a dedicated virtual machine. Shopping is treated as a first-class action: Stripe Link can create a single-use card tied to one merchant, amount and short time window, while the user approves the exact purchase.
That is a product design difference, not a clean technical lead. ChatGPT and Claude cover much of the same ground, and all three are changing quickly. Muse is simply the one most clearly shaped around handing off ordinary personal errands.
What people are handing over first
Meta has not published a breakdown of Muse tasks, so there is no defensible league table of public use cases. There is enough evidence to see an early pattern, though. Meta says buying things is one of Muse's most popular browser uses, an independent reviewer tested it, and the fight with Amazon made shopping the clearest breakout example.
Shopping begins before checkout. Muse can research products, compare options, look for a sale and ask a follow-up such as size or budget. A TechRadar reviewer used it to find shorts and received a shortlist before Muse requested an Amazon login.
But that exact route has already changed. On 21 September, Amazon blocked Muse from shopping on its store, and users saw a message saying access by an unauthorised AI agent violated Amazon's conditions. Muse may still shop elsewhere, but “use Muse to shop Amazon” is no longer reliable advice.
The next cluster is saving money and undoing chores. Public posts collected by Messaging Agents include users claiming that Muse switched insurance, negotiated bills and found money they were owed. These are self-reported stories, not audited results. They show the appeal of the use case: the agent has to search, compare, contact a company and finish a process, rather than merely explain what the customer could do.
Inbox and calendar work is less dramatic and probably easier to trust. In the same TechRadar test, Muse found a utility-company email and prepared a reply for approval. Meta's examples include following school messages, moving dates into a family calendar and turning a saved recipe into a grocery list.
Muse can watch rather than act: monitor things you care about, track a goal, and return only when something changes. This may become its most useful habit because the person does not have to remember to ask again. For now, that is a core product promise, not a measured public favourite.
The useful version needs access
Muse becomes more helpful as it can see more of your life. Email gives it dates and receipts. A calendar gives it time. Shopping accounts give it addresses and order history. Those connections also increase the damage a mistake or a malicious instruction hidden on a webpage could cause. Meta itself says Muse will make mistakes and that prompt injection remains an open problem.
Learn the product by increasing access in steps. Start with public research. Then try a read-only connection. Keep sending, booking and buying behind one-time approvals until Muse has earned your trust on small, reversible jobs.
Meta says Muse conversations and data inside its virtual machine are not shared with its advertising systems, and users can opt out of model training. Its security post also notes that a shop may treat Muse's visit as the user's activity, which can still influence ads outside Muse.
Muse is an early answer to a larger question: what should happen after an AI gives good advice? Its answer is to keep the context, do the clicking and come back at the decision point. The appealing part is the finished errand. The hard part is deciding how much of your digital life the errand is worth.