Meta Muse: 5 Surprising Features Worth Trying

Calling an AI agent the easiest ever sounds like faint praise, but the host's argument is that the simplicity is deceptive. According to Meta's official Muse page (ai.meta.com), the system is a personal agent running inside a secure virtual computer with its own browser, powered by the Muse Spark model built for real-world agent work. The phone and desktop apps being nearly identical is the first hook: full power everywhere you go.

Everything revolves around one main chat. Side chats exist for exceptions, but the default flow stays in a single thread, which removes the mental overhead of remembering where each task lives. The host says dictation is his primary input, especially on the phone, and that single-threaded design is what lowers the barrier for first-time agent users.

Single-chat design: main flow, side chats, ideas and goals

The ideas tab turns past conversations into suggestions for what to try next. After asking about climbing gyms in Los Angeles, the agent offered to build a day-by-day climbing schedule for the trip. The goals tab works the same way in reverse, generating checkable items from what you discuss, and this very video ended up as a goal. Neither tab is mandatory, but both make the agent feel like it remembers you.

The first standout skill is personal shopping. Muse browses with its own browser, finds listings, and summarizes them up to the point of purchase. The host hunts a camera first found on Facebook Marketplace, then asks for eBay equivalents; the agent returns seven listings mixing used, refurbished, buy-it-now and auction formats. Behind that flow sits the Link wallet for agents: Stripe reports (stripe.com) over 1 million businesses accepting instant checkout, a consumer network above 300 million, and a single-use virtual card scoped to each approved purchase.

The most reassuring part is the double approval architecture around money. When a login is needed, Muse never asks for the password directly; it surfaces a secure login card and keeps credentials in a vault. According to the Meta Help Center (meta.com), every purchase requires explicit approval, Link keeps the real card number private, and orders are fulfilled under the merchant's own terms. Even a restaurant without a stored card can be paid through Link, but the final word always stays with the user.

Personal shopper: browser, secure login and Link payments

The second skill is the voice memo. A long answer comparing AI models collapses into a spoken summary within seconds of asking. Better still, a side chat can be instructed to attach a voice memo to every reply from then on, which turns walks and housework into hands-free conversation time. For a phone-first user this small detail quickly becomes a habit.

One level up, the same research task becomes a full 6.5-minute podcast episode with chapter-style controls for skipping and speed. From there the host connects the dots to morning briefing automation: a recurring task that researches trending consumer AI tools every morning and delivers them as audio. That is contextual automation at its simplest, set once and heard daily.

From voice to voice: memos, podcasts and the morning briefing

The fourth skill is phone calls, tested here with early access. The scope is narrow: official US businesses only, customer-side errands such as stock checks or reservations, never personal calls or cold outreach. Every call starts with a written plan showing who is called, the number, the speaking persona, and preferences like never using the last name. Approval happens twice, the call itself cannot be listened to live, and the outcome arrives as a written record. Checking a book at Barnes and Noble is the canonical example.

The honesty test is built into the video. Shopkeepers often realize the voice is synthetic, ask whether it is an AI, hear yes, and hang up or point back to the website. The host has the agent call his own company and records it: the caller introduces itself as Brett, confirms it is an AI assistant calling on Paul Lipsky's behalf, asks only for operating hours, and closes on the 24/7 answer. PYMNTS reports (pymnts.com) that Meta briefly tested human contractors for these calls after merchants hung up, that some employees warned about leaks and weakened security claims, and that spokesperson Daniel Roberts called feedback overwhelmingly positive while promising proper disclosures before any wide rollout.

Phone calls: the plan, the double check and the shopkeepers' reaction

The fifth skill is file access, and it needs the desktop app. The host keeps a screenless Mac Mini running around the clock, opens a document on its desktop with a single command, gets a summary, and has that summary prepended to the file. Compared with the remote-computer flow of ChatGPT and Claude, which takes several menu steps, Muse answers from inside any chat. TechCrunch notes (techcrunch.com) the Mac app landed on September 18 with opt-in access to files, messages, calendar, notes and mail, always asking before sensitive actions, a point Alexandr Wang underlined in his announcement.

The bonus is price. The weekly usage quota barely moves; the host sits at 15 percent with two days left before reset. An invite code grants both sides 1 billion extra units, redeemable up to 30 times, and his settings show a 30 billion balance. Top5Apps calculates (top5apps.ai) the free tier at 100 million units a week, Power at 20 dollars a month for 500 million, Maximum at 100 dollars for 3 billion, with the 1 billion invite bonus being one-time, roughly ten weeks of free allowance, and the Muse unit being a metered consumption unit rather than a model token. The closing verdict stays balanced: not the most powerful agent, but simple and generous enough to try, with an eye already on Grockbot for the next step.