Meta Muse has surged up US app charts since its September 2026 launch. As Meta plans AI glasses and the pocket-sized Muse Charm, this article examines what the agent can do, where service access limits it, and how it compares with ChatGPT Agent, Claude Cowork, Gemini Spark and Grok Bot.
Meta’s new personal AI agent, Muse, has had a fast start. Less than three weeks after its September 8 launch, the app had reached the top of the US free-app charts on both Apple’s App Store and Google Play. At Connect 2026, Meta set out its next ambition: making Muse accessible through AI glasses, a pocket-sized device called Muse Charm, and more shopping and productivity services.
The interesting question is what happens after someone asks an agent to do a task. Muse can continue working when its app is closed, but finishing a job also depends on permission to use personal data, access to outside services, and the ability to recover when something goes wrong. Those questions matter across a growing field of personal AI agents, not just at Meta.
How Quickly Has Meta Muse Grown?
Meta introduced Muse on September 8 in the United States through iOS, Android, the web, and WhatsApp. A Mac app followed. The company describes Muse as an agent that can work across connected services: it can research options, fill forms, help arrange travel, and prepare or send emails. For longer jobs, it can keep working after a user leaves the app and return when it needs approval.
The early download figures are substantial, although they are estimates rather than a measure of continuing use. Sensor Tower estimated more than 3.4 million downloads as of September 24. Other firms reported different totals: Appfigures estimated roughly 2.3 million installs on the same date. Muse reached No. 1 on the US App Store on September 18 and on Google Play on September 19, according to the same reporting. The app was available in the US and Canada by late September, following its initial US rollout.
Muse’s appeal is easy to understand. A user can give it a goal, leave it to handle intermediate steps, and check back when a decision is required. Meta says it can remember preferences and offer relevant suggestions without being prompted each time. The basic version is free; Reuters reported subscription plans at $20 and $100 a month for heavier use. But a chart position tells us how many people were curious enough to download it; it does not yet show how reliably Muse completes tasks or how many people will keep using it.
What Meta Announced for Muse at Connect 2026
Meta’s Connect announcements put Muse at the centre of its AI glasses plans. In the coming months, wearers are expected to be able to call their agent by name and ask it to act on what they are looking at, such as a product on a shelf or a flier on a wall. This visual assistance depends on camera-equipped glasses; it should not be confused with Meta’s new camera-free audio glasses.
Those Ray-Ban Meta Audio glasses start at $349 in the US, are available for preorder, and are scheduled to ship on October 13. They offer hands-free audio and access to an AI assistant without a camera. Meta also presented longer voice conversations with Muse and Muse Realtime Avatar, technology intended to give an agent a more expressive on-screen presence. These were announcements of an expanding experience, not proof that every demonstrated feature is already available to every user.
Muse Charm takes a different approach. Meta previewed a small device intended to let people speak to their agent without first opening a phone app or putting on glasses. Meta’s written recap describes a pocketable device with real-time voice interaction. Reporting on the demonstration describes a screen displaying an avatar and a fingerprint sensor for starting a conversation. Bloomberg’s reporting adds front and rear cameras and a 5G modem. Zuckerberg said Meta was targeting December 2026 for shipment, but the final price and some product details had not been settled. Treat Charm as a preview with a target launch window, rather than a device consumers can buy today.
Meta separately announced VR glasses for spring 2027, priced at $1,299.99 in the US. The company has described their entertainment and workspace features in detail; it has not made an equally specific announcement about what Muse will do on that device.
The Services Muse Needs to Get Work Done
An agent can make a good plan and still fail at the final step if it cannot reach the relevant service. Meta says Muse launched with dozens of partners and access to Shopify’s catalogue. At launch, it supported checkout through Stripe’s Link; Meta said Shop Pay was coming soon. At Connect, it named further retail partners, including Walmart, Best Buy, Gap, Sephora, Ulta, and Wayfair, as well as PayPal and work integrations. Expedia was described as coming soon. The announcement is a roadmap as well as an account of existing connections, so availability should be checked service by service.
Meta also announced that Muse would get its own email address. A Mac update added computer use: with permission, Muse can operate applications on a user’s Mac while a task continues. This gives it another route to complete work that a direct service connector may not cover.
Yet permission from a user is only one side of the equation. Amazon blocked Muse from shopping on its platform, citing concerns about an unauthorized agent’s access. The dispute illustrates a constraint for every cross-service assistant: its reach depends both on what a user authorizes and on what outside platforms allow. More connectors could make Muse more useful, but their reliability and terms of access will matter as much as the number of partner logos.
How Muse Compares With Other Personal AI Agents
Muse has arrived in a competitive market. Its most distinctive proposition is the combination Meta is assembling around it: a continuing agent, access through messaging, planned use on glasses, and a dedicated pocket device. Other agents already offer meaningful forms of tool use and background work, often through different interfaces.
ChatGPT Agent brings research and action into a widely used chat interface. It can browse websites, draw on connected sources, work with files, and carry out multi-step tasks. OpenAI says it seeks explicit confirmation before actions with real-world consequences, such as purchases. For someone whose work starts with a question or a document, that familiar conversation is a practical access point. OpenAI’s product description also acknowledges risks from malicious instructions embedded in sites an agent visits.
Claude Cowork is oriented toward completing work with documents, files, desktop tools, and connected data. That makes it particularly relevant to people handing an agent a research, analysis, or production task rather than asking for occasional help while moving through the day. Anthropic describes Cowork’s workflow capabilities and its work within office tools and enterprise data; its appeal should not be reduced to writing quality alone.
Gemini Spark draws on Google’s ecosystem. Google describes it as a personal agent that can continue working in the background and connect with Workspace tools such as Gmail, Docs, and Sheets. It also says Spark should ask before high-stakes actions such as spending money or sending an email. Google announced an expansion to Pro subscribers in India in July 2026, making it an especially relevant comparison for Indian readers. Persistent work is therefore a feature Muse shares with a competitor, not a category Muse created on its own.
Grok Bot takes the idea of continuing work further through agents that have their own cloud computers. According to xAI, they can sign into tools, work across apps, and return when approval is needed. Its framing is closer to assigning a job to a digital teammate. Whether that suits a person depends on how much access they are willing to grant and how they want to supervise work in progress.
The wider field includes Instinct, which describes proactive help informed by connected messages and device context, and specialist assistants such as Lindy for routine scheduling and inbox work. Microsoft 365 Copilot is closely tied to work in Microsoft’s applications. Technically inclined users can also explore systems such as OpenClaw and Hermes Agent, where setup and control differ from a ready-to-use consumer app. These are different approaches to personal assistance, with different demands on the user.
The comparison is more useful than a single ranking. Where does an agent live: in a chat, an office suite, a cloud computer, or a wearable? Can it keep working after a session ends? Which services can it actually use, and when must it seek approval? Those questions are likely to determine everyday usefulness more than an impressive demonstration alone.
What Still Has to Work: Reliability, Permission, and Trust
Meta says each Muse runs in a dedicated cloud-based virtual machine. It describes a separate monitoring agent, user-controlled service access, approval before sensitive actions, and an audit trail. These are important design commitments. They also need to hold up during ordinary, messy tasks involving personal accounts and third-party websites.
Reuters reported that some internal testers encountered reliability failures and inappropriate exposure of personal data before launch. It also reported on a test in which human contractors handled some calls associated with Muse; Meta said that test was rolled back while it worked on safeguards and disclosure. Neither report establishes that every user will encounter those problems. Both show why the way an agent completes a task can matter as much as the finished result.
Muse’s early downloads show strong curiosity, while its expanding hardware plans show how seriously Meta is pursuing an agent that people can reach throughout the day. The harder measure will come later: whether it finishes useful tasks consistently, makes its actions understandable, respects permission boundaries, and can work with the services people actually depend on. That is the test facing Muse, ChatGPT, Claude, Gemini Spark, Grok Bot, and the rest of the personal-agent field.