The best thing about Meta’s new AI agent might be what it forces everyone else to do.
Meta launched Muse on September 8, promising an assistant that can take on everyday tasks and keep working after you close the app. It runs on its own cloud computer, uses a browser, and connects to services you already use.
The pitch: give it something to accomplish, then let it handle more of the work.
My take? It's a really good tool that will only improve as they continue to push their models farther, and rest assured, they're still pushing hard.
Also, more companies building credible alternatives at or near the frontier is a win for consumers. Meta and xAI continuing to fight their way up the leaderboard gives OpenAI and Anthropic more reasons to improve their products, justify their prices, and fix the things users keep complaining about.
There’s increasingly less room for mediocrity. Lovely.
Muse still has to earn its place in people’s lives. A launch announcement can establish ambition; everyday use will establish whether it deserves access to your inbox.
What Muse actually does
Muse is the personal agent. Muse Spark is the model powering it.
That distinction matters: the product announced this week packages AI into something people can delegate work to. Meta’s examples include booking travel, filling out forms, negotiating bills, and turning saved Instagram recipes into grocery lists. You can communicate through the Muse app or WhatsApp, and it can notify you when it needs approval. Those are Meta’s advertised capabilities, rather than results from our own testing.
The underlying technology has been advancing separately. Meta released Muse Spark 1.3 on September 2, emphasizing improvements in coding, tool use, and sustained work across multiple steps.
For anyone still sorting out the terminology, our beginner’s guide to AI agents explains the basic ingredients: a goal, useful context, access to tools, and boundaries around what the system can do.
Muse’s appeal is putting those ingredients within reach of someone who has absolutely no interest in spending Saturday configuring an agent.
The early reaction: excitement, with a trust problem attached
There is some encouraging early experience to point to. Journalist Alex Heath, who used Muse before launch, described himself as impressed and called it “OpenClaw for normies.” His interview with Mark Zuckerberg frames accessibility and security as central to the product. (FYI, this is an excellent interview, and worth your time if you want to know the mindset behind Meta's AI moved and data center approach. -CN)
That is a useful early signal. It is also one person’s experience over several days, not a verdict on long-term reliability.
Axios’s launch coverage asks whether Muse’s usefulness will justify giving it greater access to personal data.
WIRED similarly foregrounds trust, examining the security promises behind an assistant designed to act across people’s digital lives.
These questions belong together. An agent becomes more useful when it can reach the information and services needed to finish a task. Every additional connection also raises the stakes when it misunderstands something.
A chatbot giving you a bad restaurant recommendation is annoying. An agent making the wrong reservation adds a cancellation policy.
How to access Muse and get started
Muse is initially available to U.S. adults ages 18 and older, according to the Associated Press. Meta says it is rolling out on iOS, Android, and the web, with AI glasses support coming later.
Start at Muse’s official website or use the product link in Meta’s announcement.
The practical setup below follows published documentation; we couldn’t verify the live signup screens.
- Complete the account setup. TechCrunch reports that signup requires a payment card, including for free access.
- Connect one service to start. Choose something relevant to your first task. Meta documents controls over connected services and their permissions, including whether email access allows reading or sending.
- Review the training setting. Meta says eligible interaction data is used for model training by default, with an opt-out switch in Muse settings.
- Give it a specific, bounded job. Explain the outcome, relevant constraints, and what you want to approve.
- Check the result before expanding its responsibilities. Start with work you can easily assess, then add connections as they become useful.
For example, a sensible first request would be:
Review my calendar for next week and suggest three 45-minute blocks for focused work. Explain your choices. Don’t create or move any events yet.
That is a suggested starting prompt, not a task we tested. It gives you a way to judge whether Muse understands your schedule before allowing it to change anything.
A shopping task could follow the same pattern: specify the item, budget, delivery deadline, and requirement to approve the final purchase. Clear instructions make the result easier to evaluate.
What Muse costs
The launch options are:
- Free: $0 per month.
- Power: $20 per month.
- Maximum: $100 per month.
The paid plan names and prices are reported by TechCrunch. Heath’s launch interview reports up to 100 million tokens per week on the free tier.
Tokens are the units models process while working. That allowance doesn’t translate neatly into a fixed number of completed tasks: a lengthy assignment can involve repeated reading, planning, and tool use.
We couldn’t verify exact paid-tier allowances from a live pricing page. Check the subscription details presented in your account before upgrading.
My recommendation is to start free. Give Muse a few recurring jobs, see whether it completes them well, and let actual usage determine whether a subscription makes sense. Buying the most expensive plan before finding a useful task is a very modern way to procrastinate.
Read the privacy promises carefully
Meta’s security documentation describes a separate system called Sentinel that controls permissions and communication outside the agent’s environment. Credentials are stored separately from the main agent, and purchases require approval.
Those protections matter. They also come with an important distinction: the launch version does not prevent Meta from accessing data when necessary to operate, support, or secure the service.
A separate Muse Confidential VM, intended to cryptographically prevent Meta from accessing data inside it, is planned for later this year.
Meta also says Muse conversations and VM data aren’t shared with its advertising systems. However, activity Muse performs on websites or services can indirectly affect advertising. The company acknowledges that prompt injection remains unresolved and that Muse will make mistakes.
That makes privacy controls and permission design part of the product’s value. They deserve the same scrutiny as the quality of its answers.
Why the competition is worth welcoming
Muse arrives as other challengers are making their own pushes. Grok 4.5’s July release targeted coding, agentic tasks, and knowledge work—the kinds of jobs people increasingly expect capable AI systems to handle.
That doesn’t establish that every lab is equally strong, or that Muse has overtaken the leaders. Model benchmarks, product usability, and reliable task completion are different measures.
But consumers don’t need five companies tied for first place to benefit from competition. They need enough good alternatives that switching becomes realistic.
A capable free agent puts pressure on what paid products must deliver. Better permission controls raise expectations for everyone else. Easier setup makes a complicated onboarding process harder to excuse.
OpenAI and Anthropic should have to keep earning their place in your workflow. So should Meta, Google, and the company behind Grok.
For Muse, the next test is straightforward: can it repeatedly finish useful work with less effort than doing it yourself?
If it can, consumers get another serious option. Even people who never use it stand to benefit when their preferred AI provider has to work harder to keep them.
Related reading
- AI Agents and Automation for Total Beginners — Understand the goals, tools, context, and permissions behind products like Muse.
- Our Guide to AI Skills, Projects, Custom GPTs, Agents, Plugins, Connectors, Loops, and Codex Workflows — A practical map of the terminology surrounding customizable AI assistants.
- How to Build AI Agents for Total Beginners in 2026 — Explore how context, connections, reusable instructions, and triggers turn an agent into something useful.
