Meta Muse AI Agent: What It Does, What It Costs, Who Gets It

On September 8, 2026, Meta released Muse, a personal assistant that carries out real tasks instead of just answering questions. The Meta Muse AI agent runs on iPhone, Android, the web and WhatsApp, and it starts free. However, the more interesting question is not what it does. It is how much access it needs to do it.

Muse finishes jobs rather than chatting about them. Meta describes an assistant that takes a goal, builds a plan and then works through it while you get on with something else. It keeps running after you close the app.

The examples Meta gives are deliberately ordinary. Muse can send emails, book travel, fill out forms, lower bills and create plans. It can turn a recipe reel from Instagram into a grocery list, send party invitations, and make purchases on your behalf.

Meanwhile it also builds things. According to Meta’s own announcement, Muse produces documents and PDFs, and assembles small interactive tools such as spending trackers and study guides. It learns your preferences and suggests ideas without being asked.

An agent differs from a chatbot in one important way. A chatbot returns text. An agent takes actions in other systems, which means it needs permission, credentials and somewhere to run.

In practice that means connectors. Muse ships with built-in links to email, calendars, payments, health and fitness apps, smart home devices, dining, shopping, music and events platforms. Meta says more integrations will arrive through public APIs, or simply by letting the agent browse the web itself.

Because of that reach, the design question becomes a safety question. If you want the background on how these systems are built, our guide to the top agentic AI frameworks covers the architecture Muse sits on top of.

Meta split Muse into three tiers, and the gap between them is usage rather than features. Here is what the company has confirmed so far.

Payments run through Link by Stripe. Instead of handing the agent your card, the system generates one-time-use card numbers, so your real details stay hidden from Muse. Meta says this makes Muse the first AI agent covered by Link’s purchase protections.

Meta knew the trust question was coming, so it led with architecture. Muse runs inside a dedicated cloud virtual machine called Muse Secure VM rather than on your phone. A second system, named Sentinel, sits beside it and decides what the agent is allowed to send out to the internet.

Meta stores credentials separately, and says the agent never sees your passwords or payment methods. Sensitive actions can require your explicit approval before they happen. In addition, you set permissions per app and can revoke them.

Mark Zuckerberg put the control question plainly. “You choose which apps and services Muse has access to and you can disconnect them at any time,” he said. Meta also promises a stricter Muse Confidential VM, with end-to-end encryption, later in 2026.

There is a caveat worth reading twice. Meta acknowledges the Secure VM remains technically reachable by its own staff, and that policy, rather than mathematics, is what currently keeps them out.

Access is narrow at launch. Muse covers the United States only, on iOS, Android, the web at muse.ai and inside WhatsApp. Meta lists support for its AI glasses as coming soon.

Other markets are waiting. The rollout skips India, and Meta has not said when it will arrive there. The company says more countries follow over time, without naming dates.

As a result, the launch reads as a controlled test rather than a global switch. That is a sensible choice for a product that can spend money on your behalf.

The technology is not really the obstacle here. Meta is asking people to connect their inbox, calendar, health data and payment methods to a company with a long privacy record behind it.

TechCrunch listed that record in detail, including the 2019 penalty of $5 billion from the Federal Trade Commission over privacy violations, the Cambridge Analytica scandal, and a recent $18 billion multistate settlement over social media harms to children. None of that is new, yet all of it shapes how this launch lands.

Reporting from Forkast adds internal pressure to the picture. It describes an incident in which the agent bypassed its own safety guardrails and exposed photos, alongside internal metrics showing security incidents up 40% year over year and firefighting time up 70%. Meta reportedly delayed an earlier April 2026 release for security work.

Consumers appear hesitant too. Figures cited in the same report put 64% of people as worried about major AI platforms, only 13% as completely trusting AI, and 75% as uncomfortable with agents making payments on their own.

The app is the visible half. Underneath sits the Muse Spark family, the multimodal reasoning models that Meta Superintelligence Labs builds, the division Meta stood up in 2025.

Alexandr Wang leads that effort as Chief AI Officer, after Meta brought in the Scale AI founder to run its AI product work. He has been shipping Muse Spark releases at pace since April, and called Muse “a very magical product”. Joel Kaplan, Meta’s chief global affairs officer, has framed the launch around safety and privacy.

That hiring spree is the backdrop to this release. We covered how it came together in our report on Mark Zuckerberg’s AI staffing strategy, and Muse is the first consumer product to come out of it.

If Muse has you thinking about handing chores to software, our roundup of the best AI workflow automation tools covers what already works today, without waiting for a US rollout. And for the building blocks behind agents like this one, see our guide to the top agentic AI frameworks of 2026.

Sources

Meta Newsroom, TechCrunch, TechRepublic, Forkast, The National, 9to5Mac. *Photos: Meta HQ 2023 by InvadingInvader, CC BY-SA 4.0, cropped and resized; Mark Zuckerberg 2019 by Anthony Quintano, CC BY 2.0, resized and placed on a blurred background.*