Meta’s new Muse AI agent went live this week, and it is built to do a lot more than answer a question and disappear. Instead of a chatbot you check in on, Muse is meant to run in the background. You hand it a goal, and it keeps working the plan without you needing to open the app again.

This is the product Zuckerberg was describing a month ago in his essay on personal superintelligence, a free agent working on your behalf around the clock. Back then it was a stated direction with no shipped product attached. Now it has a name, an app, and a price tag.

The rollout started September 8, through standalone apps on iOS and Android, the web at muse.ai, or directly inside WhatsApp. For now it is limited to the United States, with Meta saying more countries will come later, no date attached. Access starts free, and two paid plans unlock heavier use: Power at $20 a month and Maximum at $100.

How the Meta Muse AI agent connects to your apps

Setting it up is a three step process. You connect the services you already use, pick a few starting tasks, then give your agent a name, a personality, and an avatar. Coverage around the launch described a wide connector list: Gmail, Google Calendar, Google Workspace, Outlook, OpenTable, Ticketmaster, Spotify, Peloton, Apple Health, Plaid for bank and card data, plus Facebook and Instagram.

I downloaded the app to see for myself. What actually showed up on the connector screen was a lot shorter. Gmail was already connected, and the “not connected” list was limited to Google Calendar, Google Contacts, Google Drive, Outlook Calendar, Outlook Email, and Outlook People. No Plaid, no OpenTable, no Peloton, no Spotify, nothing from the shopping or health side of what was announced.

Meta has a track record of rolling out connectors in phases rather than switching everything on for every account at once, so this may simply be an early stage view rather than a walked back promise. But it is a real gap between what got announced and what a real account gets on day one, and it is worth checking back on as access widens.

One connector is missing entirely from any version of the list. There is no GitHub, no developer tooling of any kind. Muse is aimed at booking dinners and tracking spending, not writing code. Meta has a completely separate product for that called Muse Code, a terminal based coding agent that has nothing to do with this launch.

Once connected, Muse can send emails, book travel, fill out forms, lower a bill, turn a saved recipe video into a grocery list, or work a longer project like planning a year of workouts or setting up a small business. Meta says it will always ask before sending a message or spending money on your behalf.

Where Muse actually runs

Does Muse keep working once you put your phone down? Yes. Each user gets their own dedicated virtual machine in Meta’s cloud, something the company calls the Muse Secure VM. It is an isolated Linux computer with its own browser, enough processing power to run smaller sub agents, and the ability to schedule jobs on its own. Your login tokens and credentials for every connected service live on that machine, not on your phone and not in Meta’s central systems, according to Meta.

So the work happens on a computer in Meta’s cloud, running the whole time, whether your phone is on the table or switched off in a drawer. A separate internal system called Sentinel watches for problems on that same machine, kept apart from Muse itself, alongside automated checks meant to catch attempts to manipulate the agent through hidden instructions in web pages or documents.

How Muse stacks up against other AI agents

Muse is not competing in the same lane as most of the agent tools getting attention this year, and the differences say a lot about who each one is built for.

Grok Bot, xAI’s own take on a persistent agent, runs on a similar idea of a standing cloud computer that keeps working on tasks over time. It launched in August behind the most expensive plans xAI and Cursor sell, but the price dropped twice within two weeks. It is now bundled into the base Cursor Pro plan at $20 a month or base SuperGrok at $30, down from the $200 to $300 range it started at.

At the other end, tools like OpenClaw or a self-hosted setup like Hermes give you full control. You choose the model, write the automations, manage your own server, and touch every part of how the agent behaves. That flexibility is the whole appeal, and it is also the whole cost. Nothing runs until you build it.

Muse sits close to Grok Bot on price now and nowhere near either of them in flexibility. Zero setup, a free tier, guardrails and approval prompts baked in everywhere. That is a fair tradeoff for someone who wants a dinner reservation made without thinking about it, and a real limitation for anyone who wants to change what the agent is actually allowed to do.

Is Meta Muse available outside the US?
Not yet. Meta has only confirmed a US launch and says other countries are coming later, without a timeline.

Does Meta Muse cost money?
There is a free tier. Heavier use requires the Power plan at $20 a month or Maximum at $100.

How is Meta Muse different from Grok Bot or a self-hosted agent like OpenClaw?
Muse trades control for convenience. It requires no setup and runs with built-in guardrails, while Grok Bot and self-hosted tools like OpenClaw give you far more control over the model and the automation itself. Grok Bot has recently dropped in price enough to sit close to Muse, but self-hosted setups still cost more in time and effort than money.

Muse is Meta’s biggest bet yet that people want an agent doing the errands, not just answering questions about them. One launch week in, with pre-release safety issues on record and a connector list still catching up to its own announcement, it still has plenty to prove.