Meta’s Muse is charting high, but the first numbers tell a messier story

Meta’s new AI agent, Muse, reached No. 2 on the U.S. App Store after launching on Tuesday. That looks strong until you look at the pace. Sensor Tower says the iOS-only app has pulled in a little over 83,000 downloads in the U.S. so far, which trails Meta’s earlier launches by a wide margin. Threads cleared 4.3 million U.S. installs on day one. Meta AI hit 108,000 in the U.S. at launch. ChatGPT moved faster still.

So yes, Muse is getting attention. It’s not a blockbuster start by Meta standards.

The chart rank is doing some heavy lifting

App Store rankings are a noisy signal. They favor speed over scale, and in a crowded category an app can climb high without reaching that many people. Muse at No. 2 tells you it has some momentum. It doesn’t tell you it has broken out.

The more useful number is the install count. 83,000 iOS downloads in the U.S. is fine for a new consumer app. It’s not explosive. And Muse is currently limited to the U.S., so the test is narrow. Meta isn’t comparing a global rollout with a domestic one, but that’s still the comparison the chart invites.

Timing matters too. Muse arrived days after Meta settled an $18 billion multistate lawsuit tied to social media harms. That won’t help with trust. Any app asking people to share personal data is going to face a higher hurdle when it comes from Meta.

That affects conversion. It affects retention. It affects how fast people are willing to connect a new assistant to their daily routines.

Meta is betting on agents that act, not just chat

Muse is part of Meta’s push into agentic AI, the current industry label for software that does work on a user’s behalf. In practice, that means going beyond answer boxes. The app is supposed to help people coordinate, plan, and possibly complete tasks with less friction than a chatbot loop.

That’s a real product shift. A chat app can stay mostly inside text. An agent needs permissions, identity, memory, integrations, and a clear line for when it’s allowed to act. That introduces engineering problems that don’t show up in demo videos.

Once an assistant starts touching email, calendars, location, contacts, or payments, the design problem changes. You need:

• strong account binding and session security
• permission scoping that doesn’t turn into a mess
• audit trails for actions taken on the user’s behalf
• reliable failure handling when APIs break
• guardrails against prompt injection and malicious inputs
• clear UX around what the agent can and can’t do

That’s where these products get judged. The demo is easy. The permissions model is the part that hurts.

Meta is also spreading Muse across iOS, web, and WhatsApp, even though Sensor Tower is only counting the iPhone install number here. That makes sense. It lowers friction and fits how people already use Meta products. It also makes the experience harder to keep consistent. A web agent, a messaging assistant, and a native app all imply different trust models. The more entry points you have, the tighter the policy layer underneath has to be.

Android is the weak spot

Muse’s Android numbers look soft so far. It sits at No. 338 in the Productivity category on Google Play. That may partly reflect the lack of public download data, but it still suggests weaker traction outside Apple’s ecosystem.

That matters because AI assistant adoption tends to follow convenience, not ideology. If the app isn’t on the device people use most, or if onboarding is clumsy, usage drops fast. Android is a huge slice of the consumer market. If Muse can’t get traction there, Meta is testing premium consumer AI on a narrower, more Apple-heavy audience than it probably wants.

Google Play rankings are also a decent stress test for whether an app has real consumer pull. An app can look polished on iOS and still flop on Android if the value proposition doesn’t hold up once it leaves one platform.

Instinct looks like the sharper comparison

The more interesting comparison isn’t ChatGPT. It’s Instinct.

Instinct is a consumer agent that works over text messages and recently hit a $2.5 billion valuation after raising $350 million. It’s moving quickly. It just added email addresses for users, and it’s building a kind of social graph for agents, where one person’s assistant can talk to a friend’s assistant to coordinate plans.

That sounds odd until you think about the tasks that make consumer AI useful: scheduling dinner, coordinating travel, exchanging details, checking availability, sharing locations, handling small jobs without pushing people back into a long app session. Text is still one of the lowest-friction interfaces for that. It’s familiar, asynchronous, and already tied to identity in a way that feels more natural than a standalone AI app asking for broad permissions.

Meta’s advantage is obvious: scale, distribution, and years of identity infrastructure. Its weakness is obvious too: trust. If you’re building an agent that needs permission to act, company history matters.

Instinct’s pitch is cleaner in one important way. It seems built around actual social coordination, not a general-purpose assistant sitting on top of a giant ad business. If agents end up relying on networks of permissioned relationships, the company that maps those relationships most naturally may have the better shot. Meta’s friend graph is huge, but it mixes close ties with passive follows and old connections. That’s useful for social media. It may be less useful for an assistant trying to figure out who really matters in someone’s day-to-day life.

The product problem is trust, not model quality

It’s easy to frame this as a model race. That misses the point. Consumer agents are going to live or die on trust, permissions, and reliability long before users care which frontier model sits underneath.

A good assistant has to make fewer mistakes than a human juggling tabs and notifications. It has to be predictable. It has to fail cleanly. And it can’t behave like a clever intern who freelances.

For Meta, that’s a hard sell. The company is asking users to let a system tied to its ecosystem know more about them so it can do more for them. On paper, that’s a reasonable trade-off. In practice, it’s the kind of trade-off people get nervous about when the company asking for it has spent years in privacy disputes, FTC settlements, and public backlash.

For engineers, this is where consumer AI gets real. The model is only one layer. The important work sits in the plumbing:

• permissions and consent flows
• data minimization
• identity resolution
• latency across tool calls
• fallback behavior when actions fail
• human override when the agent gets it wrong

Get those wrong and the app becomes an expensive demo people abandon after a few tries. Get them right and the product starts to feel useful in a way search and chat alone don’t.

Meta still has room to win

None of Muse’s early numbers mean it’s dead on arrival. A slower start can still become a real base if Meta improves onboarding, expands beyond the U.S., and finds an everyday use case people actually want. The app is already climbing the charts, and WhatsApp distribution could matter more than today’s App Store rank suggests.

But the launch has made one thing clear. Meta can’t brute-force this category the way it has sometimes done with social apps. Consumer AI agents are different. People won’t install one just because it exists. They’ll install it if it feels safe, reliable, and useful fast. That’s a harder product to ship, and a harder company problem to solve.