What Is Muse? Meta's Personal AI Agent, Explained

Every AI agent launch this cycle has reached for the same pitch: a helper that does the work rather than just describing how you might do it. Meta's version, called Muse, launched on September 8, 2026 with a line from Mark Zuckerberg that's doing a lot of rhetorical work — Muse is "the personal agent that understands your goals and works 24/7 to get things done for you." It's a big claim, and the interesting part is figuring out exactly where the claim ends and the shipped product begins.

Muse is rolling out first in the US on iOS, Android, and muse.ai, with a Mac desktop client and AI-glasses integration confirmed to be coming "in the coming months". It's free for most features, with paid tiers for heavier usage — Meta has said it plans to make money from a small transaction fee rather than a hard paywall, which matters for how you should read everything else about the product: this is built to be used by as many people as possible, as fast as possible.

What "getting things done" actually means

Strip away the marketing language and Muse's job description looks like a mix of three things: task execution, persistence, and memory.

On execution, Meta's own examples are concrete — Muse sends emails, books travel, and builds action plans for longer-term goals rather than one-shot requests. It can take a saved Instagram recipe reel and turn it into an actual grocery list and cooking plan, which is a good example of the underlying idea: most of what people save online is intent they never act on, and Muse's pitch is closing that gap. We look at that specific workflow in more depth in From Reel to Reservation.

On persistence, Muse is explicitly designed to keep working after you close the app and to come back to you when circumstances change — a flight price drops, a restaurant opens a table, a task hits a blocker that needs your input. That's a meaningfully different interaction model from a chatbot, which only exists in the moment you're typing to it.

On memory, Muse retains preferences across sessions and can make unprompted suggestions based on what it's learned about you. It can also negotiate on your behalf, which is a capability none of the other launches this cycle have claimed outright, and one worth treating with some skepticism until independent reporting shows what "negotiate" means in practice versus what it means in a press release.

The model underneath: Muse Spark

Muse runs on Muse Spark, which Meta describes as its most capable model to date, purpose-built for "real-world agentic work" rather than general chat. That framing — a model tuned specifically for taking actions across tools rather than just answering questions well — is the same bet OpenAI and xAI are making with GPT-6 Astra and Grok Bot's underlying model, respectively, and we dig into what's actually different about Spark's design in Muse Spark: The Model Behind Meta's Most Agentic Product Yet.

The part Meta is leaning on hardest: security

Because an agent that acts on your behalf is also an agent that can act against your behalf if something goes wrong, Meta built a permissions architecture around Muse before it built out the feature surface. Every Muse session runs inside a Muse Secure VM — an isolated virtual machine housing the agent and the user's data — with a separate Sentinel agent inside that same VM whose entire job is approving, blocking, or escalating every outbound network request and connected-app action Muse tries to make. Per Meta's own announcement, "nothing Muse does reaches the internet unless the Sentinel approves it," and Muse has no visibility into passwords or payment credentials at all. That's a genuinely different design philosophy from letting the agent hold your logins directly, and we cover it in detail in The Muse Secure VM and Sentinel.

The framing Meta wants you to accept is that an agent this capable can only be trusted if it's this constrained — two separate systems, one doing the work and one watching it, running in a box neither can see outside of without permission.

Where the honest gaps are

A few things Meta has stated plainly rather than left ambiguous are still limited in scope: Muse Confidential VM, which would end-to-end encrypt the entire VM so that not even Meta can access it, is described as coming "later this year" — a real commitment, but not yet shipped as of this writing. Paid tier pricing beyond "free for most features" hasn't been published in detail. And "negotiate on your behalf" is a capability claim without a documented example of what it looks like in a transcript.

How it's actually reaching people

Distribution is the other half of Meta's advantage here, separate from the product itself. Meta isn't launching Muse into a cold market the way a startup would — it's launching into an installed base of billions of people already using Instagram and Facebook daily, and Muse ships integrated with WhatsApp alongside its own dedicated app, web client, and the coming Mac desktop client. That's a meaningfully different go-to-market position than either of the other two agent launches this cycle. OpenAI's Dots reaches people through ChatGPT, Slack, and Microsoft Teams — genuine enterprise and productivity surfaces, but not consumer social platforms with billions of daily active users. xAI's Grok Bot reaches people through SuperGrok subscriptions and Cursor, both of which are opt-in tools for people already paying for AI or developer software. Meta's installed base means Muse's adoption curve doesn't depend on convincing people to download something new; the discovery step Meta had to fight to win with Facebook and Instagram in their early years already happened.

That's plausibly part of why Muse reportedly topped app charts in the US and Canada within weeks of launch — a headline number that says more about Meta's distribution muscle than about Muse's technical superiority over the other two products, and worth reading that way rather than as proof the underlying agent is meaningfully better.

None of that makes Muse's launch less real — Meta shipped a working product with a genuinely novel security architecture to a mass consumer audience in under three weeks, and it's reportedly topped app charts in the US and Canada since launch. It just means the interesting questions about Muse — how good the negotiation actually is, what the paid tiers cost, whether Sentinel holds up under adversarial use — are questions for the next six months, not this one. This series works through them one at a time, starting with the model doing the actual work.

Part 1 of 30 in the Meta Muse series. Next: Muse Spark: The Model Behind Meta's Most Agentic Product Yet. Browse more builder insights or explore AI skills at aiskill.market.