Muse is the personal AI agent Meta launched in the US on September 8, 2026 — it books travel, sends emails, fills out forms, shops, and negotiates bills on your behalf, running inside its own app, at muse.ai, or in WhatsApp. This is Muse: Meta's personal AI agent, features and capabilities, broken down after spending time inside the app and reading through Meta's own technical writeup.
Short answer: Muse is Meta's new agent that completes real tasks — emails, bookings, forms, purchases, bill negotiation — instead of just chatting. It runs on Meta's Muse Spark model inside a dedicated "Secure VM," works in a standalone app, on muse.ai, and in WhatsApp, and is free for basic use, with Power ($20/month) and Maximum ($100/month) tiers for heavier use.
Meta has been building toward this since Mark Zuckerberg started talking about "personal superintelligence" — an AI that knows your context well enough to act on it, not just respond to it. In my testing, Muse's pitch holds up better than most agent launches I've covered: it's less about a chat window and more about a task list that empties itself, with an audit trail you can actually read afterward.
What you'll need
You need a US-based account to use Muse at launch — Meta says more countries are coming, but it isn't a global rollout yet. Grab the standalone iOS or Android app, use it inside WhatsApp, or open muse.ai on desktop; all three connect to the same agent. You'll also want a payment card on file if you plan to let Muse buy anything for you (Meta routes purchases through Stripe Link, so Muse itself never sees the card number), plus a few minutes to connect the apps you actually use — email, calendar, and whatever shopping or dining app matters to you. Nothing else to install.
Step-by-step: setting up Muse
1. Download the app or open muse.ai
Muse lives in three places: a dedicated iOS/Android app, the muse.ai website, and inside WhatsApp. I started in the app since it's the most complete interface, but the WhatsApp version is the fastest way to try it if you already live in that inbox.
2. Name your agent and pick an avatar
Muse lets you name the agent and choose how it looks and talks, rather than handing you a generic assistant. It's a small touch, but it changes how the interactions read — less like querying a search box, more like messaging a person you've delegated something to.
3. Connect the apps Muse should touch
Email, calendar, payments, and category apps (health, smart home, dining, shopping, music, events) connect one at a time, and you choose which ones to grant access to. I connected email and calendar first and left payments for a second session, once I'd seen how the permission prompts behaved.
4. Give it one real task
Muse is built for concrete jobs: send this email, book this table, fill out this form, find a cheaper version of this bill. Vague requests underperform here the same way they do with any agent — see how ChatGPT’s agent mode handles the same problem if you want a side-by-side.
5. Approve sensitive actions as they come up
Before Muse sends an email, makes a purchase, or takes any action Meta classifies as sensitive, it stops and asks you to approve it. Read the specific action, not just the notification — that's true of every agent I've tested, not just this one.
6. Check the audit trail
Every run leaves a full record of what Muse did and what it's still planning to do. I check this after every task, especially the ones that ran while I wasn't watching — it's the fastest way to catch a wrong guess before it turns into a wrong booking.
Example prompts you can copy
Muse responds better to a scoped task with a clear finish line than to an open-ended ask:
- "Look at my calendar for next week and book a table for two on the night I'm free after 7pm. Ask before confirming."
- "Go through my streaming and phone bills and see if there's a cheaper plan that covers the same thing. Don't switch anything without asking me first."
- "Turn this recipe into a grocery list and check it against what's already in my saved shopping app."
- "Draft replies to the straightforward emails in my inbox from today. Save them as drafts — don't send anything."
- "Build me a four-week beginner running plan and remind me each Monday what's due that week."
Each one gives Muse a boundary — what it can decide alone and what needs your sign-off — the same discipline that matters with ChatGPT’s agents or Claude.
Common mistakes to avoid
The biggest one: connecting every app on day one instead of starting with two or three. Fewer connections means fewer permission prompts to actually read, and you learn Muse's habits faster with a small surface area. Second, skipping the audit trail — a wrong assumption buried in step three of a five-step task is easy to miss if you never open the log. Third, treating the free tier limits as a bug instead of the plan; Meta built Muse to be "free for most of what people need," and the paid tiers exist for people running it constantly, not for casual use. Fourth, handing over a task with money attached before you've watched Muse handle a low-stakes one — TechCrunch’s launch coverage is worth reading if you want to know what trust and safety questions reporters raised before you connect a card. Fifth, forgetting that Muse can be told to forget something specific it learned about you; most people never touch that control, and it's worth knowing it's there.
Muse vs. other AI agents
| Muse (Meta) | ChatGPT Agent Mode (OpenAI) | Claude (Anthropic) | |
|---|---|---|---|
| Launched | Sept 8, 2026 (US only) | Jul 2026 | Ongoing, Projects-based |
| Underlying model | Muse Spark | GPT-based agent model | Claude models |
| Where it runs | Standalone app, muse.ai, WhatsApp | ChatGPT web/app, paid plans | Claude web/app |
| Free tier | Yes, limited weekly use | No — Plus and up only | Limited, no dedicated agent mode |
| Paid tiers | Power $20/mo, Maximum $100/mo | Included in ChatGPT Plus/Pro/Team | Included in Claude Pro/Max |
| Sandbox | Dedicated "Secure VM" per user | Virtual browser | No dedicated agent sandbox |
| Can make purchases | Yes, via Stripe Link | Asks permission per action | Not built for this yet |
Muse is the only one of the three with a real free tier for agentic tasks, and the only one built specifically around a consumer's daily life — shopping, bills, appointments — rather than research or work tasks. For coding or research-heavy work, ChatGPT agent mode and Claude still have the deeper toolset.
Tools that make this easier
If you're deciding whether to trust an agent with real purchases and logins in the first place, it's worth reading what's gone wrong elsewhere — AI agents lying, cheating, and stealing is a useful reality check before you connect a payment card to anything. If you want Muse to help with actual writing — drafts, captions, longer copy — pair it with a dedicated tool from my best AI writing tools ranking rather than asking an agent built for tasks to also be your editor. And since Muse's recipe and shopping features lean on Instagram content, my breakdown of how Instagram already uses AI explains the parts of Meta's stack Muse is plugging into.
My take on Muse
Meta's own launch writeup frames it as "a personal AI agent — it doesn't just answer questions, it actually does the work," and after using it, that's the accurate way to describe it — Muse is closer to a task list with hands than a chatbot. The permission prompts and audit trail are the parts that matter most, the same way they are with any agent that can spend your money or send mail on your behalf. Whether it's better than ChatGPT's agent mode or Claude depends on what you're using it for: Muse is built for daily-life logistics, not research or code. If that's what you need done, the free tier is enough to find out before you pay for Power or Maximum.
Frequently Asked Questions
Is Muse free?
Mostly, yes. Meta offers a free tier with weekly usage limits, plus two paid tiers — Power at $20/month and Maximum at $100/month — for people who want higher limits.
How long does it take to set up Muse?
A few minutes for the basics: download the app or open muse.ai, name your agent, and connect one or two apps. Connecting everything and learning the permission prompts takes a couple of real tasks, not one sitting.
What is the easiest way to try Muse?
Message it inside WhatsApp if you already use that app daily — it's the lowest-friction way to send it a first task without installing anything new.
Does Muse see my passwords or payment details?
No. Meta says passwords and payment methods stay hidden from the agent itself; purchases route through Stripe Link, which issues one-time-use cards instead of exposing your real card number.
Is Muse available outside the US?
Not yet. It launched in the US on September 8, 2026, with Meta saying more countries will be added over time.