Meta is taking its AI ambitions a step further with Muse, a new personal AI agent designed not just to answer questions but to actually handle tasks on a person’s behalf.

Announced September 8, Muse is positioned as a more proactive kind of AI assistant. Instead of waiting for users to ask follow-up questions or provide every step of a task, Meta says Muse can take a goal, build a plan, interact with apps and websites, and continue working in the background until the job is finished or it needs the user’s approval.

That makes Muse a significant evolution of Meta’s existing AI assistant strategy. The company has been gradually moving Meta AI from a conversational chatbot toward an agent that can plan and take action, and Muse is designed to take that concept much further.

Muse is designed to do the work

The simplest way to think about Muse is as an AI assistant that can move beyond conversation.

A user can tell Muse what they want done, and the agent can determine the steps needed to accomplish it. Meta says Muse can handle relatively straightforward tasks such as sending an email or booking travel, but it can also work toward larger, longer-term goals.

For example, someone could give Muse a broader objective and have it create a personalized plan, coordinate the necessary resources and continue working without requiring constant prompts. The agent can open a browser, fill out forms and even negotiate on a user’s behalf.

Muse can also continue working after the user closes the app. If something changes or the agent reaches a point where human approval is required, it can return to the user for a decision.

That distinction is important. Traditional AI assistants are generally built around a question-and-answer interaction: you ask, the AI responds, and you decide what to do next. Meta is instead trying to create an assistant that can take responsibility for more of the workflow itself.

A personal AI that remembers context

Meta is also putting a major emphasis on personalization.

Muse is designed to remember information that matters to a person and use that context when making future suggestions. The company says the agent can act on something a user mentioned only once rather than requiring the same information to be provided repeatedly.

Meta’s examples show how this could work in everyday situations. If someone saves a recipe on Instagram, Muse could turn it into a grocery list. It could also suggest a dinner-party menu and remember guests’ dietary restrictions before helping send invitations.

The idea is to make the AI feel less like a tool that users have to operate and more like an assistant that understands the context surrounding their lives.

That also creates a much bigger privacy challenge, because an AI that can act independently needs access to considerably more information than a chatbot that simply answers questions.

Muse runs on a dedicated secure computer

To address that problem, Meta built a dedicated environment called Muse Secure VM.

The company describes it as a dedicated virtual computer in the cloud where both the Muse agent and a person’s data can reside. It includes its own browser and is designed to keep the agent isolated from other users’ agents.

Meta says a separate system-level Sentinel agent controls Muse’s ability to access the internet. Muse cannot simply perform an online action whenever it wants; Sentinel must approve the activity, with the system asking the user for permission when necessary.

The company also says Muse itself does not see a person’s passwords or payment details. Credentials are stored securely so the agent can use them without directly exposing them to Muse.

Users can decide which apps Muse can access and how much permission it receives. For example, someone can choose whether Muse can only read email or whether it can also send messages on their behalf.

Meta says users can change those permissions or disconnect services at any time. They can also tell Muse to forget specific information it has learned.

For sensitive actions, such as sending an email or making a purchase, Muse is designed to request approval. Meta also says users receive an audit trail showing what the agent has done and what it plans to do.

Muse can shop and make purchases

One of the more interesting parts of Meta’s agent strategy is giving Muse the ability to complete online purchases.

Muse can check out using Link, the payment service built by Stripe. Meta says Muse is the first AI agent covered by Link’s purchase protections, which can include coverage for eligible damaged or lost items, price drops, no-fee returns and a return guarantee.

Link’s agent wallet generates a one-time-use card, allowing the user’s actual payment card information to remain hidden during purchases.

Meta says Shop Pay will be supported in the future, while 1Password support is also coming so Muse can use existing account logins.

This is an important piece of the broader AI-agent race. An assistant that can recommend a product is useful, but an assistant that can search, compare, make a decision and complete the transaction represents a much more significant shift in how people could interact with the web.

Powered by Meta’s Muse Spark model

Behind Muse is Muse Spark, which Meta describes as its most capable model to date for this type of real-world agentic work.

Muse Spark is part of Meta’s broader AI push through Meta Superintelligence Labs. The company introduced the Muse model family earlier this year, describing Muse Spark as a model designed for reasoning, multimodal tasks and the kinds of interactions required by Meta AI.

Meta has already been moving Meta AI toward agentic capabilities. In July, the company announced features that allowed Meta AI to make plans, connect to email and calendar apps, create slides and handle tasks on a user’s behalf.

Muse represents a more ambitious version of that direction, with the agent itself becoming a dedicated personal computing environment rather than simply another feature inside a chatbot.

Privacy will be central to Muse’s success

Meta’s biggest challenge may not be whether Muse can complete tasks. It will be whether people are comfortable giving an AI agent enough access to their digital lives to make those capabilities useful.

The company is attempting to address that concern with its Secure VM architecture, permission controls, approval requirements and audit trails. Meta also says Muse does not share a person’s conversations or VM data with its advertising systems. Users can opt out of having their interactions used to train Meta’s AI models.

Meta is planning another major privacy feature later this year with Muse Confidential VM. The company says the entire virtual machine, including a user’s data and conversations with Muse, will be encrypted using a key held only by the user. That would mean even Meta could not access the contents of the VM.

For an AI agent with the ability to access email, websites, accounts and payment systems, that distinction could become one of the most important parts of the product.

Muse is rolling out in the US

Meta says Muse is rolling out in the United States on iOS, Android and the web at muse.ai. The company also plans to bring it to its AI glasses in the future.

The basic service will be free for most everyday needs, while Meta plans subscription options for people who want to do more with the agent.

Muse is therefore more than another chatbot launch. Meta is betting that the next stage of personal AI will be defined less by how well an assistant can answer a question and more by how much work it can take off a person’s plate.

If Meta can make that experience reliable while maintaining meaningful control over privacy, permissions and spending, Muse could offer an early glimpse at a very different relationship between people and their computers: one where you describe the outcome you want and let an AI figure out much of the work required to get there.