Meta's Muse puts its approval prompt on three outbound verbs

Approval is the control Meta names, so read what it covers. According to Runtimewire's reading of the Meta-authored App Store listing, Muse asks before sending messages, spending money or cancelling subscriptions, keeps an activity log, and lets a user revoke individual connectors ^[7], three verbs, all of them outbound. Neither report describes a prompt on the read side; there, the controls are the connector grant itself, revocation, and a log you read after the fact ^[7]. That gate covers three outbound verbs, guarding access to nine connected accounts ^[24].

The size of that read scope is declared, not guessed. Apple's privacy section for the app lists health and fitness records, payment and credit information, precise location, contacts, emails or text messages, browsing history and sensitive information among the data Meta may link to a user's identity, and Apple notes the disclosure comes from Meta and has not been independently verified ^[16]. Everything Muse proposes is a function of that pile.

There is also a second route to spending money. TestingCatalog describes an interactive browser-window widget, carrying a Beta label, that Muse uses to open a cinema website, select seats and handle the booking ^[15]. That account mentions no approval step inside the widget. So the thing to measure when access opens is where the gate is implemented: in the path a tool call takes out of the process, or as an instruction the model is asked to follow. Nothing in either source settles it, and Meta's own listing warns that Muse may produce inaccurate results or take unexpected actions, and tells users to monitor its work and intervene ^[8].

The rollout evidence is thinner than the feature copy. Two versions of the same TestingCatalog article, under the same headline, disagree: one puts Muse in closed beta with availability very likely restricted to Saint Kitts and Nevis and the US ^[2], the other in closed alpha, Saint Kitts and Nevis alone ^[3]. Both hedge with "very likely", which is the right amount of confidence for a market list read off a waitlist page. The dated anchors around it are firmer: Runtimewire cites a September 7th TestingCatalog post on the invite-code test ^[18], and a September 2nd Meta release of Muse Spark 1.3, a model update aimed at coding and long-running agentic work that Meta AI chief Alexandr Wang told Axios would support personal agents working continuously toward users' goals ^[19].

For the approval pattern to transfer as a standard, it has to be enforcement rather than copy, and the closed test is where that gets decided. The number worth asking for once access widens is approvals per completed task: a low count would support Meta's description of the gate as a control, while a high one would turn it into a stream of taps that users learn to clear without reading.