Meta Muse: what Meta's AI agent does, what it sees, and when in Europe

On 8 September 2026, Meta launched Muse, a personal agent that connects to your emails, calendar, payments and health apps to act on your behalf. Fifteen days later, at its Connect conference, Mark Zuckerberg made it the centrepiece of his strategy, with a video avatar, integration with smart glasses and a series of merchant partnerships.

Muse is not available in France. Yet it has a good chance of arriving there, via WhatsApp, which millions of employees already use on their phones. This guide explains what it does, how it is built, what Meta promises about your data and what remains to be verified, so you know what to think when the question arises, at home or in your organisation.

1. What Muse does

Muse functions as an agent: you give it a goal, it establishes a plan and acts in your accounts until it has achieved it, even when the application is closed.

• Target tasks: sorting and drafting emails, organising a calendar, booking a trip or restaurant, filling in forms, paying bills, making purchases.
• Connections: emails, calendar, payments, health and fitness apps, smart home, dining, shopping, music. At Connect, Meta added connectors to GitHub, Granola and Notion, as well as retailers like Best Buy, Gap, Sephora, Walmart, Wayfair and Expedia.
• Payments: they go through Stripe Link, Shop Pay and PayPal, without the agent seeing the card number.
• On computer: the Mac version, released on 19 September, works in your files, emails, messages, calendar and notes, with explicit activation and validation before any deletion or sending.
• What's coming: a video avatar with which you can dialogue in real time, integration with Meta's smart glasses, an email address specific to the agent and opening connectors to developers.

2. How it is built

Muse's architecture is more polished than that of most consumer agents. Each user has their own virtual machine, the Muse Secure VM, which contains the agent, its files and its memory.

• The virtual machine: everything Muse produces or uses for you is stored there. Meta indicates that you can inspect, modify and download these files, including the memory the agent keeps about you.
• The credentials vault: your passwords and access tokens are stored in an isolated space. The agent can use them to act, without ever reading them.
• The Sentinel: a separate component controls each network access and each action on a connector, and it is at this point that credentials are added. Each action is logged.

3. What Meta promises about your data

Meta has made specific commitments, and it is on these that Muse must be judged.

What Meta commits to doing

• Not sharing your conversations or your virtual machine data with its advertising systems, even if your account is linked to Facebook or Instagram.
• Keeping your credentials out of reach of the model.
• Letting you choose connected applications, revoke access and ask Muse to forget information.
• Offering later a Confidential VM, designed with Moxie Marlinspike, Signal's creator, in which even Meta could no longer read your data. Meta has opened the code of this version to security firms and plans a transparency log.

What remains open

• Model training. Your conversations, tool calls and exchanges between sub-agents can be used to train Meta's future models, after removing the main identifying data. This is the default setting: you must opt out yourself in the settings.
• Meta's access. As long as the Confidential VM is not available, protection relies on internal rules; encryption will come with it.
• Track record. The American press recalls FTC sanctions in 2011 and 2019 ($5 billion), the Cambridge Analytica affair and, in August 2026, an $18 billion settlement with 29 states over social media harms.
• Commercial incentive. Mark Zuckerberg announced that Meta plans to earn revenue through a small commission on transactions. An agent paid when you buy is not neutral in its recommendations.

4. Pricing and availability

Muse is free for a significant volume of use, with a quota. The American press reports two paid tiers, Power at $20 per month and Maximum at $100 per month, for intensive use; Meta has not published a complete pricing grid in its announcement.

At launch, Muse is reserved for the United States and users over 18. It is accessed via the Muse app on iOS and Android, via the muse.ai site, in WhatsApp, and on Mac since 19 September.

5. Muse, ChatGPT Work and Claude Cowork

The three major personal agents of the moment do not target exactly the same audience.

6. What about France?

Meta has given no timetable for Europe. The regulatory context is more constraining: at the end of 2025, Meta had already excluded the European Union, the United Kingdom and South Korea from its decision to use conversations with its AIs to target advertising, because personal data laws forbade it. An agent that reads emails, health data and banking information raises even heavier GDPR questions.

For a business, the issue will probably arrive through the back door. Muse is integrated into WhatsApp, already installed on many employees' phones, sometimes for professional exchanges. The day it opens in Europe, a colleague could connect their professional messaging to an agent hosted in the United States with a few clicks.

7. Questions to ask before enabling Muse

1. Which accounts to connect? Start with a secondary account, read-only when possible. Never your main email or bank on day one.
2. Is training disabled? Check the setting before the first conversation.
3. Which actions require your approval? Test a dozen tasks, including one that should trigger validation, and verify it actually arrives.
4. What happens on disconnection? Check what happens to stored data when you remove a connector or close the account, and whether export works.
5. Who else is affected? An agent that reads your emails also reads those of your correspondents. For professional use, the decision also belongs to your organisation.
6. Is the Confidential VM available? For sensitive data, wait for the encrypted version with your own key and verify it has been audited.

Frequently asked questions

Can you use Meta Muse in France? Not at the time of writing. Muse is reserved for the United States and those over 18, with no opening date announced in Europe.

Does Muse read my passwords? No according to Meta: they are stored in an isolated vault and added by Sentinel at the moment of action, without the model seeing them. However, Meta retains the technical ability to access data for support and security, as long as the Confidential VM has not been released.

Do my conversations serve advertising? Meta commits not to share them with its advertising systems. However, they can be used to train its models, by default, unless you opt out.

Is Muse free? Yes for everyday use, with a quota. Tiers at $20 and $100 per month have been reported by the press for intensive use.

What's the difference from an autonomous assistant like Hermes Agent? Muse is hosted at Meta and only works with Meta's model. An autonomous assistant is installed on your own machine and can work with a local model, without your data leaving your infrastructure. Our AI agents guide compares the two families in detail.

Going further

Muse belongs to the personal agents family, one of the four families described in our guide AI Agents in 2026: frameworks, runtimes, autonomous assistants and personal agents, with a comparison, security rules and a decision tree by profile. The nAIvigate Trust Index documents, publisher by publisher, what each model provider does with the data it receives.

Sources

• Meta, help centre: Muse privacy, safety and security
• Meta AI Research, agent security with Muse
• TechCrunch, Muse launch and trust question (8 September 2026)
• TechCrunch, Connect announcements (23 September 2026)
• MLQ, training policy and Meta access to data
• AI Weekly, Secure VM, payments and paid tiers
• TechCrunch, targeted advertising and AI conversations outside European Union (October 2025)
• Reuters via The News, ChatGPT Work launch