What is Meta Muse? Meta's personal AI agent, explained
Meta's personal AI agent on a cloud computer of its own, launched September 8, 2026: what it can reach, when it asks, what it costs, and which Muse is which.
Muse is Meta’s personal AI agent, launched on September 8, 2026. Each person’s Muse lives on its own virtual machine in Meta’s cloud, with a browser and a terminal, and works toward goals you give it: it uses the apps you connect, fills in forms, books and buys things with a one-time card number, and keeps working after you close the app, coming back when it needs your approval to send an email or make a purchase. You reach it in the Muse app on iPhone, Android and Mac, on the web at muse.ai, or in WhatsApp, if you are 18 or over in the US or Canada. It is free up to a usage limit, and the Power ($20 a month) and Maximum ($100 a month) plans raise it.
It is not Muse Spark, the Meta model it runs on, and not the Meta AI chat assistant: Meta now puts the Muse name on a family of products, sorted out below.
Which Muse? The name covers a family
- Muse, the agent (this page): the app and the computer behind it. Pre-launch reports called it Hatch, and Meta’s own security write-up confirms Hatch is “our internal name for Muse in the codebase.”
- Muse Spark: Meta Superintelligence Labs’ language models. The first launched on April 8, 2026 to power the Meta AI assistant, and Muse Spark 1.3, released September 2, is the model behind the agent. Developers can call it through Meta’s Model API and OpenRouter, so an agent you run yourself can use it too.
- Meta AI: the chat assistant in WhatsApp, Instagram, Facebook, Messenger and the Meta AI app. It runs on Muse Spark as well, but it is a separate product from the agent.
- The rest of the family: Muse Image (image generation, July 2026); Muse Glimmer, a 30-billion-parameter model with open weights under Apache 2.0 (August 10), built to run an always-on agent on a Mac or PC with one consumer GPU; Muse Code, a terminal coding agent for developers; and Muse Charm, a pocket device for talking to your Muse, announced at Meta’s Connect event on September 23 with details promised later this year.
- Not Meta’s: OpenMuse, an MIT-licensed, self-hostable personal-agent app from CopilotKit, published September 15, with goals, ideas and side chats like Muse’s and support for any agent harness. And older pages about muse.ai describe a video-hosting service that held the address until early 2026.
A computer of its own, and a guard beside it
Meta calls the machine Muse Secure VM. Its security write-up describes each one as “an isolated linux box with a browser and enough storage, CPU, and memory to do real work,” dedicated to one person: it holds the agent, your files, its memory, and the passwords and tokens for every service you connect, which stay in your VM rather than in Meta’s central systems. The agent runs in a locked-down container on that machine. Beside it runs a second agent, the Sentinel, which the first cannot override: every network request and every connector action passes through it, it decides whether to allow, refuse or ask you, and it swaps in real credentials only at the network boundary, so the agent itself only ever handles stand-ins. The design assumes the agent will sometimes be fooled by instructions hidden in what it reads (prompt injection), and Meta’s bug bounty pays up to $300,000 for valid reports, injections included.
The isolation is between users, not from Meta. The same write-up says today’s design “does not prevent Meta from accessing data when necessary to support, secure or operate the service.” A Muse Confidential VM, encrypted with a key only you hold, is promised for later in 2026. The Muse terms also let Meta “monitor, log, review, suspend, block, or modify Muse’s actions” and keep logs of what it does on your behalf.
Built on OpenClaw’s ideas
Anyone who has run OpenClaw, the open-source agent, will recognize Muse’s insides. Meta’s help page on managing your data lists files you can open and edit: Soul.md (“who your Muse is, including core truths, boundaries, and personality”), Identity.md (“its name, creature, vibe, or tag line”) and Memory.md. OpenClaw keeps SOUL.md, IDENTITY.md and MEMORY.md in its workspace, and its templates carry the same parts: Core Truths and Boundaries headings in the soul file, Creature and Vibe fields in the identity file.
When users pointed this out, Nat Friedman, head of product at Meta Superintelligence Labs, answered on X on September 21: “we built muse from scratch, but it is definitely heavily inspired as a product by openclaw,” adding that after trying OpenClaw in January he had bought hundreds of Mac minis for his team. Asked why the file names matched, he replied that “Peter got those things exactly right,” meaning Peter Steinberger, OpenClaw’s creator. The pattern underneath is the one called a cloud agent: an agent with a computer that stays on and keeps its state. What differs is who holds the machine and the model. For how the options compare, with OpenAI’s dots (announced September 29) among them, see dots, Muse or an agent you run yourself.
What it can reach, and when it asks first
Muse works through connectors, Meta’s word for the apps and services you link to it: email and calendars, and for small businesses Asana, Canva, Notion, QuickBooks, Shopify, Slack and dozens more. Retailers including Walmart, Best Buy, Sephora and Wayfair were announced at Connect, with Expedia to follow. Facebook, Instagram and Threads connect automatically if those accounts share an Accounts Center with your Muse account. For a service that isn’t listed, you can ask Muse to build a custom connector from the service’s API; Meta’s connector help page says it doesn’t review those.
Beyond connectors it browses with a Chromium browser you can watch and take over, and it pays from a wallet: at checkout, Link by Stripe issues a single-use card number tied to that merchant and amount, so neither the shop nor the agent sees your card. Meta announced Shop Pay and PayPal at Connect. Not every shop lets it in: on September 20 Amazon began refusing it with “Continued access by an unauthorized AI agent violates Amazon’s Conditions of Use,” TechCrunch reported.
On a Mac, the Muse app also works on your own computer: your files, and apps such as Messages, Notes, Reminders, Mail and Calendar. It needs macOS’s Full Disk Access and Automation permissions to do so, and each app that keeps data locally is set to off, read only, or read and interact.
How often it asks is a setting, described on Meta’s approvals help page. The default, “Ask for some actions”, asks before every write action (sending a message, making a purchase) and before important reads, and on the web asks when your information may be shared or the site is unfamiliar. “Always ask” asks before anything. Each request offers Allow once, Allow for this task, Allow for this site, Always allow or Deny, and it arrives as a card in the app, outside the conversation, with your answer going to the Sentinel rather than to the agent. A purchase needs your approval every time, showing its exact details.
What stays on the machine
Your chat history, files, artifacts, goals, reminders and scheduled tasks, the memory files and the credential store all live in the VM, and Meta says the VM is backed up continuously. You can download a copy of your chats, files and agent information from Settings, ask Muse to “forget” a person or topic, or reset it, which deletes everything and cannot be undone.
Two defaults are worth checking on the first day. Your conversations are used to train Meta’s models unless you turn off “Help improve our AI models” under Data controls; Meta removes names, email addresses, phone numbers and similar identifiers first, and the switch applies to past conversations too. And if your Muse account shares an Accounts Center with your Facebook or Instagram, the Muse privacy policy lets Meta combine information across them, except where it says otherwise: your conversations and VM data are kept from Meta’s ad systems, though a site Muse visits sees the visit as yours and may advertise to you accordingly. Signing up with an email that isn’t tied to another Meta account keeps Muse separate.
Where it runs, and what it costs
Muse launched in the US on iPhone, Android and the web; by September 29 Meta listed it for people 18 and over in the US and Canada. A Mac app followed on September 17, and Meta’s download page offers no Windows version. You can also message it in WhatsApp. At Connect, Meta announced Muse for its AI glasses “in the coming months,” a voice mode, and an email address of each Muse’s own; on September 29 it added Muse for Small Business, a set of business connectors and skills.
The plans, from Meta’s subscription help page as of late September 2026:
- Free: usage up to a limit Meta doesn’t publish; when you hit it, you upgrade or wait for it to refresh.
- Power, $20 a month: 500 million “Muse tokens” a week.
- Maximum, $100 a month: 3 billion Muse tokens a week.
The page doesn’t say how much work a Muse token buys. Meta’s longer-term revenue is meant to come from transactions: Mark Zuckerberg said at Connect that Meta expects “over time we will profit by taking a small fee from transactions,” per TechCrunch.
For developers: Muse connectors
On September 18 Zuckerberg opened connectors to outside developers: “You bring the API,” he wrote, and Muse “brings the agent, the browser, and the context of what the person actually wants.” The route is the Muse Connector Platform: you describe your product, Meta reviews the connector for “functional, security, and legal requirements” and tests it end to end, and once approved it appears in the directory inside Muse, where Meta’s editors pick which to feature. Meta has partnered with Stripe so a connector can take payments through Link.
Meta’s developer site separately takes applications for early access to connectors for the Meta AI assistant and glasses, where you describe your endpoints through a guided interface or MCP and link accounts with OAuth. Alexandr Wang, Meta’s chief AI officer, said at Connect that more than 1,500 applications had arrived in under a week. A listed connector passes Meta’s review; a custom one a user has Muse build for themselves does not.