Meta's Muse AI Agent: What It Does and What It Knows
Muse is a general personal AI agent, not an Instagram creator tool. Here is what Meta published, what an independent reviewer found, and where the privacy questions actually sit.
Verdict: misleading.
Primary source: Meta newsroom announcement of Muse, September 8, 2026 →
An agent that acts on your behalf across accounts needs credentials and memory. Whether that is private depends on design choices Meta has now described in detail.
Short answer: The framing is wrong in a way that matters. Muse is not an Instagram creator tool. Meta launched it on September 8, 2026 as a general personal AI agent that lives in its own app, in WhatsApp, on the web and on Mac, and it is designed to act on your behalf rather than to post content for you. Instagram is one of the apps it can reach. What is genuinely interesting about Muse is the security design Meta published alongside it, because that is where the privacy question actually lives.
1. What Meta actually announced
Meta’s own newsroom post is the primary document. It describes Muse as a secure, private personal AI agent that proactively helps with people’s goals and suggests ideas. The unusual part is the machine underneath: Muse runs on Muse Secure VM, a dedicated virtual machine that houses both the agent and the user’s data.
The Statesman reports that the announcement came directly from CEO Mark Zuckerberg, who described Muse as an agent that understands a user’s goals and works around the clock to complete tasks, and that it is built on the latest models developed under chief AI officer Alexandr Wang. It runs inside a chat interface, similar to the Meta AI app.
What it does is broader than the Instagram framing suggests. Moneycontrol reports that Muse can book travel, fill out forms, manage projects, send emails and help with long-term goals. Users tell Muse what to accomplish and the agent breaks the objective into smaller tasks and carries them out. Meta says it can browse websites, complete online forms, organise travel, negotiate services and keep working in the background until a task is complete. Trusted Reviews adds that Muse can answer questions, make purchases, and generate images or documents.
The model underneath is Muse Spark, which Moneycontrol reports is Meta’s most capable AI model to date, developed specifically for agentic tasks rather than chat.
2. What an independent reviewer actually found
The TechCrunch hands-on, carried by Yahoo Tech, is the closest thing here to an independent test. The reviewer instructed Muse to email a friend about a trip, and when the friend replied, had Muse package her restaurant recommendations into a polished document. In a separate task, the agent produced a list of what to pack each day for a move, with time allocations and general tips.
The reviewer’s summary judgement is that, like other AI agents, Muse is good at planning-oriented tasks that usually require some combination of searching, messaging people and booking reservations. They also found the app offers more guidance for users than ChatGPT or Gemini, which can feel like a blank canvas. The app has five tabs: the main chat, a customizable feed of the user’s interests and priorities, and an Ideas tab.
That is a real test with real results, and it does not involve content creation at all.
3. The security design, which is the actual story
This is where the privacy question sits, and Meta has been unusually specific about it.
The agent’s actions are reviewed before they reach the internet. Moneycontrol reports that a separate security system called Sentinel reviews every action before it goes out, and that users retain control over which apps Muse can access and what permissions it receives. Meta says Muse cannot view saved passwords or payment details directly. Users can disconnect services at any time, ask Muse to forget stored information, and opt out of having their conversations used for AI model training.
For actions involving purchases or sensitive communication, the AI pauses and asks for approval before proceeding.
Two further details matter. The Statesman reports that the dedicated virtual machine includes a built-in browser that stays visible to the user during tasks, which is a transparency mechanism rather than a slogan. And the same report says Meta plans to launch Muse Confidential VM later this year, which would encrypt the entire virtual machine with keys accessible only to the user.
There is also a naming collision worth recording, because it shows how a product launch goes live on platforms the company does not control. The Statesman notes that the British rock band Muse, active since 1994, holds the handle @Muse on Instagram, and that following Meta’s announcement that handle was reassigned to the new AI product, with the band’s account moved to @MuseBand instead.
4. What the business model is, so far
The Statesman reports the pricing and the commercial direction. Meta is offering a free tier for Muse, plus two paid subscription tiers, one at 20 dollars a month and the other at 100 dollars a month. Wang told Axios that most users should be able to do what they need within the free tier, and that the paid tiers help cover compute costs for heavier users. There is no advertising inside Muse right now, and Wang said the company is looking into commerce features that could bring in revenue later.
The same report says Meta plans an open source line under the same family called Muse Glimmer, aimed at running on laptops.
5. What the record does not settle
Three questions stay open, and they are the interesting ones.
The Observer frames the strategic point: trust and permission may matter as much as model performance in Meta’s vision for personal AI. Bank of America analysts, quoted by TechCrunch, put it more plainly: while feedback around Muse has been positive, privacy and trust remain key considerations for broader adoption. The Observer also notes the precedent: Muse’s rise in the app stores comes after the company’s smart glasses stirred controversy and privacy concerns.
Rollout geography is limited. The Statesman and Moneycontrol both report that Muse is currently available only in the United States, on iOS, Android and the web, with Meta’s AI glasses to follow. That means most readers of this article cannot test it, and any claim about how it behaves in the wild is currently untested outside the United States.
And the memory question is unresolved by design. Moneycontrol reports that Muse remembers preferences and details shared during conversations in order to suggest actions proactively, with the example of converting an Instagram recipe into a grocery list or recommending a dinner menu while remembering guests’ dietary restrictions. A system that remembers preferences in order to act without being asked is the privacy question in miniature. The answer the people who built it prefer is to put the agent in a dedicated VM with user-only keys and a visible browser. Whether that is enough is a question for the users of the product, and for the regulators who will decide what counts as consent.
For how the same trade-off looks when the model runs on your own hardware instead of someone else’s cloud, see our Ollama versus LM Studio comparison and our local model and VTuber piece.
Sources to check yourself
- Meta newsroom, September 8: the announcement itself, the definition of a personal AI agent and the Muse Secure VM.
- Moneycontrol, September 9: the task list, Muse Spark, Sentinel, the approval pauses, the user controls and Muse Confidential VM.
- The Statesman, September 10: the visible browser, the free and paid tiers, the US-only rollout, Muse Glimmer and the Instagram handle reassignment.
- Trusted Reviews, September 24: the surfaces Muse runs on and what it can be asked to do.
- TechCrunch via Yahoo Tech: the hands-on test with the email and packing tasks, the app’s five tabs and the Bank of America analysts on privacy.
- Observer: why permission and trust matter more than model performance in Meta’s strategy.
Frequently asked questions
Q: Is Muse an Instagram feature? A: No. Meta launched it as a personal AI agent available in its own app, WhatsApp, on the web and on Mac. Instagram is one of the Meta apps it can connect to.
Q: Does it create content for creators? A: That is not what Meta published it for. The documented tasks are booking, forms, travel, email, documents and planning. A hands-on review found it good at planning-oriented tasks rather than content production.
Q: Can it spend my money without asking? A: No. Meta says the agent pauses for user approval on purchases and sensitive communications, and that Sentinel reviews every action before it reaches the internet.
Q: Where is it available? A: The United States only, on iOS, Android and the web, according to The Statesman and Moneycontrol, with Meta’s AI glasses to follow.
Reader discussion
Conversation for this story opens with memberships. Until then: found an error? Report it — we correct visibly.


