On 8 September, Meta launched Muse, a "personal AI agent" it says can put AI in the hands of virtually anyone. The assistant handles everyday tasks and projects — online shopping, sending emails, planning a trip — and, once given a goal, Meta says it can work on its own, opening a browser, filling out forms, and even negotiating on users' behalf.
Related: Earlier coverage: Reporting indicates Meta's Muse agent enters closed alpha with invite codes.
Muse rolls out in the US on iOS, Android, and muse.ai, with AI glasses "coming soon." It will be free for most users, though Meta said there will be paid subscriptions "for people who want to do more." The company did not specify the nature and costs of those paid plans, nor outline what limits free users face.
For lengthy tasks, Meta says Muse keeps working in the background after users close the app, returning when something changes or when it needs approval for an action such as making a purchase. "Muse is a first step: an agent that takes on more of the work so people can focus on what matters to them," Meta said.
Confirmed
Muse is powered by Meta's in-house Muse Spark model, the first model out of Meta Superintelligence Labs. Meta is positioning the agent as the centerpiece of its AI strategy going forward, after a multi-billion-dollar overhaul aimed at catching up to OpenAI, Anthropic, and Google.
Meta pitches ease of use as the differentiator. "Anyone can use it out of the box, no technical experience required," the company claims, adding that "there's no learning curve." Users interact through the Muse app or WhatsApp, and Meta says the agent will remember details shared even once, enabling unprompted suggestions.
On privacy, Meta says users can opt out of their interactions being used to train its AI models — it is not known whether that setting is on or off by default — and can instruct Muse to "forget" specific things it has learned. Muse runs on a virtual computer in the cloud, kept separate from other users' agents. Meta says it has no visibility into passwords or payment methods, and a second agent, Sentinel, patrols the same virtual machine so nothing Muse does reaches the internet without approval.
Meta said it plans a more secure "confidential" version of that virtual machine later this year, encrypted so that even Meta cannot access it. 1Password and Shop Pay support are on the way, and Muse can already check out securely with Stripe's Link.
The agent arrives into a crowded field. OpenAI's ChatGPT Work, Anthropic's Claude Cowork, Microsoft's Copilot Tasks, SpaceXAI's Grok Bot, and the viral open-source agent Moltbot are mostly aimed at businesses or technical users, though Google's recently announced Gemini Spark appears to target a similar consumer audience. Meta calls Muse "the world's first personal AI agent built for everyone," a claim the competitive lineup does not support on its face.
Unknown
Pricing is the largest gap. Meta confirmed paid tiers exist but gave no price, no feature split, and no usage limits for the free version. That makes it impossible to judge how far "free for most users" extends before the subscription wall appears.
Meta's performance claims for Muse Spark remain vendor-reported. According to Meta's model evaluations, the model's "Contemplating" mode surpasses Gemini 3.1 DeepThink and GPT 5.4 Pro on the benchmark known as Humanity's Last Exam — a figure that has not been independently replicated in the material available here. Meta says Muse Spark can orchestrate teams of agents in parallel without affecting response time, and that it partnered with more than 1,000 physicians on health training data for "more factual and comprehensive responses."
Trust is the other open question. Meta has endured a trial over the harms of social media and repeated privacy ordeals, including the Cambridge Analytica scandal. Its AI efforts have produced a "Discover" feature that displayed other users' prompts, conversations, and image outputs, a support chatbot that helped hackers take over more than 20,000 Instagram accounts, and a short-lived AI-generated clickbait news feed. Whether Muse's sandboxed virtual machine answers those concerns will only be testable once the product is in wide use.
Our take
Meta's real bet is distribution, not capability. Muse Spark is a compact model, and Meta's own framing puts it behind frontier rivals on raw performance. What Meta has is WhatsApp, Instagram, and a billion-plus consumers who will never open a developer portal. That advantage is real — and it is why the permission model matters more than the benchmark table.
