Muse AI is Meta’s answer to the question the whole tech industry has been circling for two years: what comes after the chatbot? Launched on September 8, 2026, Muse is a personal AI agent that doesn’t just answer your questions — it does things for you. It sends emails, books travel, fills out web forms, negotiates on your behalf, and even buys things for you, quietly working in the background after you’ve closed the app.
One thing to clear up first, because it trips a lot of people: the muse.ai web address used to belong to something completely different — an AI video-hosting platform, a searchable library of AI videos that review sites were writing about as far back as 2021. Meta has since taken over the domain for its new agent. So if you search “muse.ai review” and land on a video platform, you’re looking at the old tenant. This guide is about Meta’s Muse: what it is, what it can do, what it costs, what’s new, and whether you should trust it with your digital life.
The short version: Muse runs on Meta’s Muse Spark models inside a private cloud computer assigned to you, it’s free for most people with $20 and $100 monthly tiers for heavier use, and it’s currently available in the US only. The full version — features, pricing details, the security design, the controversies, and how it compares with ChatGPT, Claude, and Gemini — is below.
What is Muse AI?
Muse is a personal AI agent built by Meta — specifically by Meta Superintelligence Labs, the division led by Chief AI Officer Alexandr Wang, the former Scale AI chief executive. Mark Zuckerberg announced it himself on Instagram on September 8, 2026, framing it as a step toward what he calls “personal superintelligence”: an AI that knows you, works for you, and handles the tedious parts of your digital life.
The distinction that matters most is the one between a chatbot and an agent. A chatbot answers when you ask. An agent acts. You tell Muse to book a trip, and it doesn’t give you a list of flight websites — it opens the browser, compares options, and books. You tell it to lower your bills, and it goes and negotiates. That shift — from advice to action — is why Meta, OpenAI, Anthropic, and Google are all racing to ship agents at the same time.
Muse lives in a few places. There’s a standalone app for iOS and Android, a web version at muse.ai, and it works inside WhatsApp as a chat contact. Support for Meta’s AI glasses is planned. Right now it’s available in the United States only, and you have to be 18 or older to use it. Meta hasn’t given a timeline for other countries.
Under the hood, the agent runs on Muse Spark, Meta’s flagship family of agentic AI models — more on those in the updates section below. It remembers your preferences across conversations (mention a dietary restriction once and it recalls it when sending party invitations later), it keeps working after you close the app, and it asks for your approval before doing anything sensitive. You can also name your agent, pick an avatar, and dial its proactivity up or down.
Muse AI features: what it can actually do
Launch coverage and early reviews agree on the headline: Muse is built for doing, not chatting. Here’s what that covers in practice.
Everyday tasks, end to end. Sending emails, booking travel, filling out web forms, negotiating — Meta’s own example is haggling for a higher price when selling your car. One demo that got attention: save an Instagram recipe reel, and Muse turns it into a grocery list. Send party invitations, and it remembers which guests are vegetarian.
It works while you’re away. Close the app and Muse keeps going in its cloud workspace. It comes back to you when something changes or when it needs your approval — a flight price dropped, a form needs a signature, a purchase needs a yes.
Goals become plans. Tell it something long-term — a yearlong exercise plan, starting a small business — and it builds a personalized plan, coordinates your time and resources, and monitors progress over weeks and months.
A real browser, file system, and terminal. Inside its cloud workspace, Muse opens an actual browser, navigates real websites, works with files, and can even build small tools for you. If a service you use has a public API, Muse can connect to it with credentials you provide. If there’s no API, it just uses the browser like you would.
It makes things. Documents, PDFs, web pages, interactive artifacts — spending trackers, study guides, that kind of thing. Think of it as a junior assistant with a laptop that never sleeps.
It can actually buy things. This is the feature that separates the demos from the real product. Purchases go through Link by Stripe: Muse pays with one-time-use virtual cards, so your real card details are never exposed to merchants. Meta says Muse is the first AI agent covered by Link’s purchase protections — damaged or lost items, eligible price drops, returns. Shop Pay and 1Password integrations are coming next.
App connectors, opt-in one at a time. Muse ships with built-in connectors for common services, and you grant access per app — email read-only versus send, for example — and disconnect whenever you want.
A usage meter keeps it honest. An in-app gauge shows how much of your usage is left, and warns you before the free allowance runs out, offering the paid tiers. One catch worth knowing: you have to add a payment card to start, even on the free tier.
How Muse works: your Secure VM and the Sentinel watchdog
Meta’s pitch for Muse rests heavily on its security architecture, and it’s worth understanding because it’s genuinely different from how most AI assistants are built.
Every user gets a dedicated cloud computer — Meta calls it the Muse Secure VM. The agent itself lives there, along with your data, your credentials, and its own browser. It’s your private workspace in the cloud, separate from every other user’s.
Then there’s Sentinel, a separate system that controls everything leaving that VM. Before Muse takes an action, Sentinel matches it against an authorized policy. If the action fits, it goes through. If it doesn’t, you get an approval prompt. Meta’s VP Tarek Sheasha put it bluntly: Muse “can’t override” Sentinel. It’s a hard boundary, not a suggestion.
A few more design choices worth noting. The agent never sees your real credentials — connector logins are kept outside its reach — and it never shows you your own passwords or card numbers either. There’s an audit trail, so you can see what Muse did and what it plans to do. You can opt out of having your data used for training, and Meta says Muse conversations and VM data aren’t shared with its ad systems. There’s even a public bug bounty of up to $300,000 for qualifying Muse security reports.
Here’s the honest caveat, and it’s an important one. Meta acknowledges that, technically, it could access your VM today — policy forbids it, but the capability exists. That’s according to Meta VP David Singleton, speaking to WIRED. The fix is already on the roadmap: a Confidential VM planned for later in 2026, which encrypts the whole workspace with a key only you hold, so that even Meta can’t see inside. Until that ships, you’re trusting Meta’s policy, not its math.
Muse AI pricing: free vs Power vs Maximum
Is Muse free? For most people, yes — with an asterisk. Here’s the lineup as of September 2026:
Free — $0. Meta says this covers “most of what people need.” Usage is metered, with the in-app gauge showing what you have left. The asterisk: you still have to add a payment card to sign up, even if you never pay.
Power — $20/month. More agentic usage for everyday tasks, aimed at people who lean on the agent heavily.
Maximum — $100/month. The highest usage allowance, for the heaviest users.
What Meta won’t say is exactly where the lines are. The company declined to share specific usage limits, saying they may vary by “country, surface and system conditions.” Executives have set expectations low on purpose: Alexandr Wang said “the vast majority of users should be able to do what they need within the free tier.”
A word of caution about numbers you’ll see floating around. Some third-party guides claim specific token allowances for each tier — tens or hundreds of millions of tokens per week, different prices on web versus in-app. Those figures come from single, unverified sources and appear in no primary Meta material. Treat them as rumor, not fact.
One more pricing signal: on September 15, Meta expanded its subscription push with new AI-focused plans spanning consumers, businesses, and creators — a sign that Muse’s tiers are part of a bigger plan to make AI subscriptions a real revenue line for the company.
Latest Muse updates: Spark 1.3, Voice Transcribe, Glimmer & more
Confusingly, “Muse” is two things: the agent product, and Muse Spark, the family of AI models that powers it. The agent launched September 8, but the models have been arriving all year. Here’s the timeline that matters:
April 2026 — Muse Spark 1.0. The first model from Meta Superintelligence Labs. A natively multimodal reasoning model — and notably proprietary, a break from Meta’s open-source Llama tradition. It started as a private API preview.
July 9, 2026 — Muse Spark 1.1. Public preview on the Meta Model API with a 1-million-token context window. Priced at $1.25 per million input tokens and $4.25 per million output tokens, with $20 in free credits for new API users.
Early August 2026 — Muse Spark 1.2 and Muse Code. Alongside the model came Muse Code, a beta coding tool for macOS and Linux with persistent coding agents, parallel sub-agents working in separate worktrees, and a replay-exact event log.
Around August 10, 2026 — Muse Glimmer 30B. Meta’s return to open source: an Apache 2.0 licensed model with roughly 30 billion parameters, optimized for agents, that runs on a 24GB VRAM GPU. Developers noticed.
September 1–2, 2026 — Muse Voice Transcribe and Muse Spark 1.3. Voice Transcribe is a real-time multilingual speech model — trained on 70+ languages, validated on 25 — priced at $3 per 1,000 audio-minutes, and it powers Mac dictation via the Fn key. Spark 1.3, the current flagship, uses roughly 20% fewer tool calls and 25% fewer tokens than 1.2 at the same API pricing, and scored 62 on the Artificial Analysis Intelligence Index.
September 8, 2026 — the Muse agent launches on iOS, Android, the web, and WhatsApp. September 15 brings the subscription expansion. If you’re keeping score, Meta shipped a model, a coding tool, an open-source release, a speech model, and a consumer agent in about five months — the pace is the story as much as any single release.
Muse AI vs ChatGPT vs Claude vs Gemini
Muse didn’t launch into a vacuum — it landed in the middle of an agent war. Every major lab shipped its own agent this year, and they genuinely do different jobs. Here’s the honest map:
ChatGPT’s “Work” mode (launched July 9) is the deliverables machine. It pulls context from Slack, Gmail, Drive, and Salesforce, and works for hours on research-heavy output. Since September 3, Plus subscribers and up also get GPT-6 Astra. If your work lives in documents and data, this is the one reviewers reach for.
Claude’s “Cowork” (generally available since April 9) is the computer-control specialist. It reads folders you grant it, works across your applications, got its own browser in a side panel in late August, and runs scheduled Routines. File-heavy, system-level work is its home turf.
Gemini is the Workspace native. Deep Gmail and Google Workspace integration is the whole pitch, and Google released Gemini 3.8 Flash in the first week of September. If your life runs on Google apps, the integration is hard to beat.
Muse, by contrast, is aimed at your personal life rather than your work life. One reviewer’s shorthand has stuck: everyday errands and shopping go to Muse, research and diverse work go to ChatGPT, computer control goes to Claude, and Gmail/Workspace automation goes to Gemini. That’s a simplification, but it’s a useful one — Muse is the only one of the four whose launch demo centered on negotiating your car sale and turning a recipe reel into groceries.
If you’re still weighing the broader field of assistants, our roundup of AI tools similar to ChatGPT covers more of the landscape.
The trust question
Here’s the part Meta can’t market its way around. An agent that reads your email, holds your credentials, and spends your money is asking for a level of trust no chatbot ever needed — and it’s asking from a company with a trust deficit.
The facts, fairly stated. On the architecture side, the Secure VM plus Sentinel design is serious engineering: per-user isolation, a hard policy boundary the agent can’t override, credentials kept out of the agent’s reach, an audit trail, granular per-app permissions you can revoke anytime, training opt-out, no ad-system sharing, and a $300,000 bug bounty. That’s a stronger starting position than most consumer AI products.
On the other side of the ledger: Meta itself acknowledges it could technically access your VM today — the Confidential VM that would make that impossible is still “later this year.” Reuters reported that during internal testing, employees ran into reliability problems and incidents where sensitive data was exposed or transmitted without authorization. The original April launch was reportedly delayed for extra security work. And the timing didn’t help: Muse arrived less than two weeks after Meta’s $18 billion multistate settlement over social-media harms. The public mood was captured in one widely quoted line: “Love the idea and the polish. Hate the owners.”
Outside experts add a structural warning worth hearing. When a chatbot gets something wrong, you get a bad answer. When an agent gets something wrong, the blast radius is bigger — reputational damage from a misfired email, leaked medical information, a drained bank account. That’s CDT’s Miranda Bogen, and she’s describing the category, not just Meta’s version of it.
None of this means Muse is unsafe to try. It means the right posture is eyes-open: start with low-stakes tasks, keep purchase approvals on, review the audit trail, and decide for yourself whether Meta’s policy promises are enough until the Confidential VM ships.
What this means practically
If you’re in the US and 18 or older, you can download the app or open muse.ai today, add a card (yes, even for free), and hand it a chore. Start small — have it summarize your inbox, plan a weekend trip, or turn a saved recipe into a shopping list — and watch how it handles approvals before you let it near anything with money attached.
The bigger picture is that the agent era has properly begun. Siri’s long-promised AI overhaul is expected next year, ChatGPT and Claude already act on your behalf, and now Meta wants the job of running your personal life. Whether Muse wins that job will come down less to its feature list — which is genuinely strong — and more to the trust question above. Meta is betting that convenience wins. The next twelve months will show whether it’s right.
New to the whole category? Our guide on how to use AI agents walks through the basics of working with agentic AI safely.
FAQ
What is Muse AI?
Muse AI is Meta’s personal AI agent, available at muse.ai and as an iOS/Android app and WhatsApp contact. Unlike a chatbot that only answers questions, it performs tasks for you — sending emails, booking travel, filling out forms, negotiating, and making purchases — working in the background after you close the app.
Is Muse AI free?
Yes, there’s a free tier that Meta says covers “most of what people need,” with metered usage shown in an in-app gauge. You do have to add a payment card to sign up, even on the free tier. Heavier users can upgrade to Power ($20/month) or Maximum ($100/month).
How do I sign up for Muse?
Download the Muse app on iOS or Android, visit muse.ai on the web, or message it in WhatsApp. You’ll need to be in the United States, 18 or older, and have a payment card ready for signup — even for the free tier.
Is Muse available outside the US?
No. Muse is currently US-only and 18+. Meta hasn’t announced a timeline for other countries.
Can I trust Muse with my passwords and bank details?
Meta’s design keeps real credentials outside the agent’s reach and uses one-time virtual cards (via Link by Stripe) for purchases, so merchants never see your card. The caveat: Meta acknowledges it could technically access your cloud workspace today — policy forbids it, but the capability exists until the planned Confidential VM (user-held encryption key) ships later in 2026. Start with low-stakes tasks and keep purchase approvals on.
What’s the difference between Muse and Muse Spark?
Muse is the agent product — the app and assistant that does tasks for you. Muse Spark is the family of AI models that powers it (currently version 1.3), also available to developers via API. Think of Spark as the engine and Muse as the car.
References
- TechCrunch — Meta debuts its Muse AI agent. Will consumers trust it? — September 8, 2026
- Fast Company — Meta wants its new Muse AI agent to run your digital life — September 9, 2026
- The Wall Street Journal — Meta Launches a Personal AI Agent Designed to Be Easy to Use — September 9, 2026
- SiliconANGLE — Meta debuts its ‘secure by design’ personal AI agent Muse, stock rises 6% — September 8, 2026
- TechCrunch — Meta expands subscription push with new AI-focused plans — September 15, 2026
- Motley Fool — Mark Zuckerberg’s Meta Is Charging Consumers for a Personal AI Agent for the First Time. Here’s What $20 and $100 a Month Could Add to Revenue — September 11, 2026
- TechNode — Meta launches Muse personal AI agent in the US — September 9, 2026
- SiliconANGLE — Meta says it has caught up with Anthropic and OpenAI with Muse Spark 1.3, its most powerful LLM yet — September 2, 2026
- VentureBeat — Meta returns to open source with Muse Glimmer, an Apache 2.0 licensed 30B parameter AI model optimized for agents — August 10, 2026
- TechCrunch — Meta launches Muse Code, an AI agent for large code bases — August 5, 2026
- 9to5Mac — Meta launches Muse Voice Transcribe for real-time voice dictation on Mac — September 1, 2026
- Reuters — Meta debuts Muse Spark 1.1 with preview open to developers — July 9, 2026
Leave a Reply