Meta Muse Explained: Features, Pricing, Privacy and Where It Works

Muse is Meta's personal AI agent: an assistant that does not just answer questions but acts for you, sending email, booking travel, filling in forms and buying things. It launched in the United States on September 8, 2026 and, by late October, had reached Canada, Japan and Singapore. In its first eight weeks it also picked up a long list of security and privacy incidents. This guide pulls the whole story into one place: what Muse does, what it costs, how its security is meant to work, what has gone wrong so far, and how to decide whether to give it access to your accounts.

What Muse actually does

Meta's launch announcement describes Muse as an agent that keeps working after you close the app and comes back when something changes or it needs your approval. In practice that means it can:

• Handle email and forms: draft and send replies, and fill in web forms for you
• Plan and book: research and book travel, and turn long-term goals into step-by-step plans it works through on its own
• Shop: buy things through Stripe's Link, using one-time card numbers. Meta says Muse is the first agent covered by Link's purchase protections
• Turn content into tasks: for example, make a grocery list from an Instagram recipe reel and remember your guests' dietary restrictions
• Make suggestions on its own, based on what it has learned about you over time

You connect services one at a time (email, calendar, payments, smart home, dining, shopping) and choose how much access each one gets, for example read-only email versus permission to send. If a service has a public API but no built-in connector, Muse can set one up using credentials you provide. If there is no API at all, it falls back to using a browser. Muse runs on iOS, Android, the web at muse.ai and inside WhatsApp chats.

The models behind it

Muse launched on Muse Spark, Meta's agent model family. The version current at launch, Spark 1.3, was released on September 2 and is tuned for long, multi-step tasks. Developers can use it through Meta's Model API, but its weights are closed, so you cannot download or self-host it. On October 12, Meta added Muse Wonder and the larger Muse Wonder Full, built around what Meta calls "Active Memory" and "Active Focus" (CNBC has the details). For model-level pricing and benchmarks, see our Muse Wonder explainer.

What Muse costs

Axios reported that Muse shows no ads and that Meta earns money from subscriptions and transaction fees instead. You need a payment card to sign up, even for the free tier. Singapore prices come from Techgoondu's launch coverage, which only gives the S$22 to S$139 range. The per-plan figures in the table are the two ends of that range, not separately confirmed prices for each plan. If you want extra usage on any plan, our Muse referral code guide explains how the token bonus works.

How the security is supposed to work

An agent that can spend money and send messages is a far bigger target than a chatbot. Meta's design has five main parts:

• Muse Secure VM: every user's agent runs in its own isolated cloud virtual machine.
• Sentinel: a second agent on the same machine, walled off from Muse. Meta says nothing Muse does reaches the internet unless Sentinel approves it, and Sentinel asks you before sensitive actions.
• Hidden payment details: Muse never sees your passwords or card numbers. Purchases go through one-time Link cards.
• No ad targeting: conversations and VM data are not shared with Meta's ad systems.
• Your controls: an audit trail of what Muse has done and plans to do, a "forget" command, and an opt-out from training. Training on your queries is on by default, according to Axios.

Meta has also promised a Confidential VM, encrypted with a key only you hold, later in 2026. As of October 30 it has not shipped, so today Meta can still technically access the data in your VM.

What has gone wrong so far: a timeline

Muse's first two months show how much a broadly permissioned agent can expose. Here is what happened, in date order:

None of these incidents breaks the core design. Sentinel and the per-user VM were not bypassed. But each one shows a weak point the design does not cover: your own device, the sharing features, the human-assisted parts of the service, and what Meta itself chooses to remember. Those weak points are what to judge, more than the VM diagram.

Hardware: glasses, Charm and DIY gadgets

At Connect 2026, Meta said Muse is coming to its AI glasses "in the coming months". It also showed Muse Charm, a small pendant for talking to the agent, due around December (TechCrunch lists everything announced). For tinkerers, Meta open-sourced the Muse Gadget SDK on October 2. It includes ESP32 firmware and a Linux SDK under the Apache 2.0 licence, and lets boards like a Raspberry Pi 5 or a Home Assistant setup talk to Muse. The device code is open, but it only works with Meta's service and needs a Muse API token, so treat that token like a password. Our Muse glasses explainer covers the hardware in detail.

Is Muse available in India and other countries?

Muse is officially available in four markets: the US, Canada, Japan and Singapore. Meta has announced no date or pricing for India, the UK or the EU. The web app may load elsewhere, but signing up still needs a supported region and payment card, so be wary of "Muse is now live in India" claims until Meta confirms them on its own newsroom. For connected payments in India, compare Muse with how AI agents are being wired into UPI, which takes a very different approach to approving each payment.

Our take: should you connect Muse to your accounts?

This assessment is based on Meta's published documentation and the independent reporting linked above. Muse is not offered where our editor is based, so we have not tested it hands-on and will update this section when we can. Weighing all of it, we suggest a staged approach:

• Fine to try: research, trip planning, recipe and shopping lists, and drafting emails that you send yourself. Give read-only access where you can.
• Only with approvals on: purchases through Link, and calendar changes. Check the audit trail weekly for the first month.
• Hold off for now: work email, health or financial documents, and anything about other people you would not want profiled. Wait for the Confidential VM and an independent audit of it.
• Never: share Muse chats publicly, put a gadget API token in a public code repository, or run Muse on a computer where you install unvetted software.

The bigger lesson applies to every agent, not just Meta's. An agent's safety depends on the weakest system it touches, so give it the least access that still makes it useful. The same principle runs through our coverage of Meta's earlier Hatch agent tests.

Sources

• Meta Newsroom: Introducing Muse (September 8, 2026)
• Axios: Meta debuts Muse
• TechCrunch: Everything new coming to Muse
• TIME: Muse privacy investigation
• Ars Technica: shared-conversation exposure
• Techgoondu: Muse goes live in Singapore

Frequently asked questions

Related articles

Comments