Muse AI Explained: Meta’s Personal Agent — Features, Pricing, Updates & More

What would you hand to an AI if you actually trusted it to finish the job? Not answer a question about the job. Finish it. The emails sent, the tickets bought, the forms filled. That’s the bet behind Muse AI, Meta’s new personal agent, and it’s a stranger bet than it looks, because the technology mostly works and the trust part is still an open question.

What Muse AI Actually Is

A chatbot waits for your question. An agent takes your goal. If you’re asking what is Muse AI, that distinction is the whole answer: Muse AI lives on the agent side of it. You connect it to your email, calendar, payments, and the apps you use for shopping, travel, health, and the smart home, then you hand it errands. Meta’s own description of the pitch is plain: once a person shares a goal, Muse helps build a plan, organizes time and resources, and carries on the work by itself. It opens a browser. It fills out forms. It negotiates. Search trends have already shortened the name to “Meta Muse”, same product, either name finds it.

Muse AI is Meta’s personal agent that books, buys, and emails for you. Meta’s own examples are deliberately mundane: selling a car for more, getting a bill lowered, adjusting a training plan as life shifts. The product is aimed at people with no technical background, the interface is a chat window, inside a dedicated app or directly in WhatsApp.

Under the hood is Muse Spark, Meta’s model built for agentic work. The assistant runs on iOS and Android, on the web, inside WhatsApp chats, and, as of a recent addition, on Mac. If the agent-versus-chatbot distinction still feels fuzzy, our complete guide to AI agents breaks down what “autonomous” actually means in practice.

The Machinery: A Private Computer and a Second Agent Watching It

Every Muse user gets a dedicated virtual machine in Meta’s cloud, the company calls it the Muse Secure VM, with its own browser, file system, and storage. That’s where the agent lives and where the credentials for every connected service are kept. Close the app and the work continues; the little computer doesn’t.

Then there’s Sentinel. A separate agent runs on that same machine, isolated from Muse at the system level, and nothing Muse does reaches the internet unless Sentinel approves it. Sensitive moves, sending an email, completing a purchase, pause and wait for the person’s explicit approval. Muse itself never sees your passwords or card numbers; credentials go into secure storage it can use without viewing, including passwords typed directly into its browser.

Two more details: the audit trail logs everything the agent has done and plans to do, and memory is editable, anything learned can be reviewed or forgotten. Missing for now: the promised Confidential VM, encrypted with a user-held key so not even Meta could access it. Until it ships, the privacy separation is policy, not cryptography.

The Errands It Runs, and the Leash It Wears

So what do people actually ask it to do? The real-world tests from its first weeks sketch a consistent picture. Moving quotes from local companies. Booking a date night through a reservation site. Cleaning thousands of promotional emails out of an inbox. Scouring theater sites for the cheapest decent seat to a show. The common thread isn’t glamour, it’s tedium. Multi-step coordination with waiting in the middle: email someone, wait for the reply, act on it.

The leash has three strands: approval gates before sending or spending anything (reviewers confirm they fire), scoped permissions you can revoke at any time, and the audit trail logging every action. One practical consequence reviewers kept noting: the agent is only as good as the access you grant it. Decline to connect your accounts and it degrades into a research helper that hands you links.

Payments run through Link by Stripe, which generates a one-time-use card for each purchase so the merchant never sees the real card details, the first AI agent covered by Link’s purchase protections. One caveat with teeth: Amazon has blocked the agent from its site, calling it unauthorized, so the “buy anything anywhere” pitch already has a large exception.

Muse AI Pricing, in Tokens Rather Than Dollars

Start with the odd part, because Muse AI pricing confuses everyone: Meta prices Muse AI in tokens, not tasks. A token is roughly a word-fragment of computation, and agentic work burns through them fast, every browser scroll, form fill, and background check draws from the pool. The plans don’t unlock extra features. They buy bigger pools.

A few honest footnotes. Meta hasn’t published what any token pool converts to in finished tasks, so cost-per-errand math is guesswork. Subscriptions renew monthly, require a Meta account, and buyers must be 18 or older in a supported region. One reviewer’s data point: weeks of regular use, including heavy shopping sessions, consumed only about a tenth of the free weekly allotment.

The Numbers That Moved a Market

The launch landed harder than most product debuts. Apptopia data put Muse at 2.8 million downloads in its first 12 days; on a like-for-like basis, that’s 1.8 million against 1.3 million for ChatGPT’s own first 12 days. Meta’s shares surged more than 20 percent in the weeks after launch, adding over $200 billion in market value. JPMorgan’s analysts wrote that Muse has the potential to become the most widely used consumer AI application since ChatGPT. Bank of America pointed to what’s resonating: the agent’s “ability to execute multi-step tasks and workflows with limited user intervention.” Jefferies went further, sketching a billion users by end of 2027 with 3% paying, about $10.8 billion in annualized revenue, and set an $875 price target. (I’ve watched this movie before: analyst scenarios are maps of enthusiasm, not promises.)

The app topped Apple’s App Store charts after the Connect showcase. None of this proves the product works. It proves the market thinks personal agents might finally be a consumer category.

Three Weeks of Real Testing: What Held Up

Launch demos lie; the useful evidence is what happened when reviewers handed Muse AI real errands, and the first three weeks produced an unusually thorough public test record.

Reading through the launch-week test accounts, the one that stuck with me was Barron’s Josh Schafer’s, partly because he approached Muse AI as a skeptic and refused to connect Gmail or Calendar at all. Even hobbled, the agent earned its keep: it analyzed a credit-card statement PDF and flagged a forgotten Adobe subscription, saving nearly $25 a month, then walked him through the cancellation. It found cheaper lookalike pants, flight options for a wedding, and three nearby doctors taking his insurance for a foot problem he’d been putting off for months. His summary of the trade: “Google is for finding things; I’m for getting things done”, Schafer’s “I” meaning himself with the agent running the errand. (Honestly, that one line explains the product better than Meta’s whole launch keynote.)

Other testers’ results rhyme: Yahoo Finance’s Daniel Howley got moving quotes from four companies, two calling back within minutes. CNN’s Lisa Eadicicco got a date night booked and a packing plan built. The Verge’s Emma Roth watched it clear thousands of promo emails and flag items already sitting in a shopping cart before buying only what she asked for. Then the failures, just as instructive: it sent testers to restaurants that had been closed for years, admitted to inventing a phone number, and picked “no tip” on a bakery order without asking. Coordination tasks: strong. Judgment calls and fresh facts: verify everything.

The Setup That Keeps You in Charge

None of the testers who had a good time winged it. The ones who stayed in control followed roughly the same playbook: connect the calendar first, low risk, high usefulness, and connect email as read-only to start. On a Mac, check two permissions most people never look at: Full Disk Access in system settings, and the Messages connector inside the app.

Two settings deserve attention on day one. Training contribution, letting Meta use interactions to improve its models, is on by default; the switch lives under Settings, then Data controls. And purchases should stay on one-at-a-time approval, with the full order read before confirming, tip line included.

That caution isn’t theoretical. The first three weeks produced three trust incidents worth knowing: Inc.’s Jason Aten found message history syncing after he’d declined Messages access (Meta called the integration opt-in and owned the bad explanation); Ars Technica’s Patrick Wardle found a macOS voice-transcription flaw redirecting audio, which Meta hotfixed within about 12 hours; and Reuters reported human contractors had quietly handled some Muse phone calls during testing, a practice Meta rolled back after internal criticism. None of them sank the product. Together, they’re why the checklist above exists.

Good first errands for a Muse AI agent: get quotes from three movers with explicit instructions not to book anything; ask for a shortlist of restaurants for a specific evening with a “show me before booking” rule. Tedious, checkable, low-stakes. That’s the sweet spot. For a broader walkthrough of working with agents safely, see how to use AI agents.

Announced vs. Actually Shipping

Here’s the honest ledger of what’s live versus merely announced.

The default avatar is a cream-colored character called Jolly, customizable down to its name and clothing; Zuckerberg’s own agent is a toga-wearing Agrippa. The glasses demo was the shakiest moment: the agent kept mistaking side conversation for commands. Treat the whole roadmap as directional until each item actually ships, I’ve learned to read launch roadmaps as wish lists with good graphic design.

Three Mix-Ups Worth Clearing Up

First, for everyone still asking what is Muse AI versus Meta AI, Muse is not the new name for Meta AI. Meta AI, the assistant inside Instagram, Messenger, and WhatsApp, is a separate product. Muse is a standalone agent with its own app, its own virtual machine, its own memory.

Second, it can’t spend your money on a whim. The approval gate before purchases and outgoing emails is architectural, not a suggestion, and every hands-on tester confirmed the checkpoint fires. What it can do is surprise you with judgment calls inside an approved task, like the no-tip bakery order. Approve the details, not just the total.

Third, the free tier isn’t unlimited and the privacy story has two chapters. Free usage runs on a weekly meter that refreshes. And while Meta doesn’t share conversations with its ad systems, training use is a separate toggle that starts on, plus the Confidential VM that would make Meta itself blind to your data doesn’t exist yet. If privacy is the blocker, that’s the feature to wait for.

Should You Try Muse AI?

Here’s the dry version: the Muse AI agent is a genuinely capable errand-runner with serious architecture behind it, free to try, carrying real privacy trade-offs and a mixed trust record at three weeks old. If you live in a supported region, the cost of finding out is an afternoon and a Meta account.

Apple is running the same playbook from the other side, Siri AI, its rebuilt assistant, is rolling out with a similar personal-context pitch.

The review that stuck with me while researching this wasn’t a headline, it was Claire Vo’s detailed teardown in Lenny’s Newsletter, the one people who build agents for a living kept passing around. What impressed her was the craft: a permission pattern that asks exactly when access becomes relevant and re-checks before acting, an activity feed logging every tool call, and a family morning brief that arrived better than versions she’d coaxed out of rival agents. Then she listed the browser-shopping failures anyway. That’s the product, as far as I can tell from three weeks of other people’s testing: unusually thoughtful scaffolding around an agent that still invents phone numbers. Start with read-only access, keep approvals on, verify the facts it hands you.

For the full Muse AI vs ChatGPT breakdown and where Meta Muse sits next to the chatbots you already know, see AI tools similar to ChatGPT. The roadmap items above will change this assessment as they ship.

FAQ