The consent step is the visible part
Meta introduced Muse this week as a personal agent for users in the United States, and the list of what it plugs into is the point: email, calendars, payments, health and fitness apps, the smart home, dining, shopping and music. Services are connected one at a time, and the agent completes purchases through Link by Stripe. The product runs on Meta's own Muse Spark model and reaches users at muse.ai, in iOS and Android apps and inside WhatsApp chats. It is free until usage passes the free tier, at which point Power costs 20 dollars a month and Maximum 100 dollars a month, which is why a payment card is required before the first errand.[1]
Two paths matter more than the connector list. When a service has no built-in connector but offers a public API, Muse sets up the connection using credentials the user hands over. When there is no API at all, it reaches the service through the browser, the way a person would. Meta says the agent keeps working after the user leaves the app and learns from conversations what matters to them. Each of those sentences describes a flow: something the user holds moves into the agent, and something the agent observes stays with it.[1]
What the subscriber could not obtain
A month earlier the same question arrived from the other end. Anthropic told affected customers it had become aware of a bad actor using common infostealer malware to steal Claude login sessions and then spending subscribers' token allowances through them. Grant De Swardt, an independent AI consultant in East Sussex, noticed unexplained consumption on his account on 4 August. Anthropic suspended the compromised accounts, invalidated sessions and server-side tokens, issued partial refunds and emailed a malware warning. It declined to provide itemised usage details or guidance on how a subscriber could recognise misuse. Across the cases described on Reddit and GitHub, between 10 percent and 100 percent of an allowance went without matching activity.[2]
Put the two side by side and the same missing artefact appears. The consent screen in Muse documents the moment a connection is made; it does not produce an itemised list of what the agent afterwards did through that connection. Subscribers to Claude needed exactly that and did not get it: without an itemised account of the consumption, a customer cannot line the provider's figure up against their own activity and show which part was theirs. A consent screen answers who may act; an itemised list answers what was done.[1], [2]
The document that would close the gap
This is the shape of the argument made here on 6 September about Deliveroo riders, who could see the fall in their own ledgers while the pay calculation behind each order stayed hidden. The rider held one half of the arithmetic and the platform held the other. A Muse user is offered the same division: granting the connection is theirs to do, and the work done through it sits on the side of the ledger they cannot open.[1], [3]
What would settle it is narrow and nameable. A Muse activity view, or a support page describing one, that lists each action taken through a connected service with its timestamp and outcome would let a user compare the agent's work against their own calendar entries and card statements. Until such a page exists, the honest reading is that the announcement covers the connection and stops there. That leaves missing information rather than evidence, and it is the piece worth asking for first.[1]