Meta introduced Muse, a personal AI agent product, on September 8. It’s led by chief AI officer Alexandr Wang. The product takes the form of a chat interface that, according to Meta, is more proactive and runs for longer stretches than a typical chatbot.
Where It Runs
Muse runs on a dedicated virtual machine in Meta’s cloud, complete with a browser the user can watch. Anything that has to be clicked through on a webpage — sending an email, filling out a form, booking a trip — happens inside that browser. Users can name their agent, give it an avatar, and adjust how it talks. The underlying model is called Muse Spark.
The rollout starts in the US, launching simultaneously on iOS, Android, and the web. Support for Meta’s smart glasses is listed as “coming soon.”
Three Pricing Tiers
Free, $20 a month, and $100 a month. Wang’s line on pricing: “For the vast majority of users, they should be able to accomplish what they need within the free tier.” His long-term framing for the product line is building personal superintelligence that helps people reach their goals and pursue what they care about.
Meta hasn’t disclosed what separates the tiers in practice — how many tasks the free tier can run per day, how virtual-machine runtime is metered — or how many users signed up at launch.
The Privacy Layer
Muse’s network and service access is governed by a separate system called Sentinel, which decides what the agent can and can’t touch. Compared with Meta’s existing AI products, Muse adds a toggle that can block queries from being used for other purposes.
There’s also a confidential mode, where the user’s virtual workspace is invisible to Meta. It’s currently listed as “planned,” with no rollout date. Until it ships, there’s no public information on how activity logs inside the workspace are retained, for how long, or who can access them.
An agent that can open its own browser, log itself in, and fill out its own forms carries a different order of risk than a chatbot that just answers questions. Meta calling out Sentinel as a distinct piece at least signals that this layer got separate design attention.
What It Means for China’s Own Agent Race
The initial rollout excludes mainland China, and glasses support is even further out. That doesn’t mean the region lacks equivalents: over the past year, Chinese phone makers and several large-model companies have been building assistants that carry out tasks across apps, generally along two paths.
One mirrors Muse: put the agent in a cloud virtual machine and let it open a browser there. That approach doesn’t tax the local device and lets failed attempts restart cleanly, but login sessions and activity traces have to pass through the cloud. The other keeps everything on the phone, using system accessibility permissions or a manufacturer’s own APIs to tap through the screen — data never leaves the device, but the agent can only operate apps already installed on that one phone, and switching devices means starting over.
Neither path has solved the same open question: when an agent hits “confirm payment” or “submit” on a user’s behalf and something goes wrong, whose fault is it? Muse’s public materials don’t address that either.
Sources: Axios, CNBC, Slashdot compilation, CocoLoop; pricing tiers, launch markets, and the Sentinel mechanism follow Meta’s public statements, and the comparison with alternatives in China is editorial analysis.