Meta launches Muse personal AI agent in U.S. amid privacy scrutiny

Muse, developed by Meta Superintelligence Labs and powered by the company’s Muse Spark foundation models, works by connecting to users’ email, calendar, payments, health, shopping and other apps, according to Meta. Once a user gives the agent a goal, Meta says, Muse can work independently — opening a browser, filling out forms, negotiating prices and continuing background tasks after the user closes the app. It can also turn Instagram and Facebook recipe videos into grocery lists, compare prices and place orders.

Muse is available in the U.S. starting Tuesday on iOS, Android and the web at muse.ai, as well as through WhatsApp. Support for Meta’s AI glasses is coming soon. The agent is free to start, but Meta is introducing two paid subscription tiers: Power at $20 per month and Maximum at $100 per month. Alexandr Wang, Meta’s AI chief, told Axios that most users should be able to do what they need within the free tier, and that paid tiers help cover computing costs. Purchases made through Muse use Stripe’s Link payment infrastructure, which gives the agent a single-use card number and offers purchase protections, according to the company. Meta is exploring taking a cut of agent-related commerce transactions, Wang told CNBC, but no concrete plan has been settled.

The company is trying to distinguish Muse through ease of use and security. “Anyone can use it out of the box, no technical experience required,” Meta said in a blog post, adding that there is no learning curve. Users interact with Muse through a chat interface and can customize their agent’s name and avatar. The agent will check in before high-risk actions such as sending email or completing a purchase, but can handle low-risk tasks without approval.

Meta says Muse runs on a dedicated virtual machine in its cloud, called Secure VM, with its own browser and an AI monitor named Sentinel. In a technical post, Meta said Muse has no visibility into users’ passwords or payment details, and every interaction with the outside world goes through Sentinel, which the agent cannot override. Users can opt out of having their conversations with Muse used to train Meta’s AI models; if they do not opt out, the company says it removes critical personally identifying information first. Meta also says it does not share Muse conversations or data with its advertising systems. A confidential version of the virtual machine, encrypted so Meta itself cannot access it, is planned for later this year.

The launch comes at a fraught moment for Meta. The company recently agreed to a multistate settlement over claims that it misrepresented harms on Facebook and Instagram; TechCrunch reported the agreement at $18 billion, while CNBC put the figure at nearly $17 billion. Meta also faces personal injury and school district lawsuits over similar allegations. At the same time, Wall Street is pressing the company to show returns on its heavy AI and data-center spending, and regulators and the public are increasingly concerned about AI agents’ cybersecurity risks.

Chief Executive Mark Zuckerberg has called personal agents the foundation for Meta’s next wave of products and revenue lines. Wang told CNBC that Muse was designed so it “feels very approachable and friendly and explainable, and it doesn’t feel too complicated,” while behind the scenes it may be doing advanced coding workflows and integrations. Zuckerberg hired Wang last year as part of a $14 billion investment in Scale AI, according to CNBC.

Muse faces competition from similar agentic tools offered by OpenAI, Anthropic, Google and startups. Meta acknowledges the agent will make mistakes. “Muse can and will still make mistakes, but we expect they’ll be much less frequent and cause much less damage due to the safety systems we’ve built in,” the company said. Meta is inviting third-party security researchers to find vulnerabilities in Muse through a bug-bounty program, and says it has designed the agent with privacy controls that go beyond its earlier AI products.