APoV 001: Muse, Consumer AI Agents, and the AI Super Gateway

Over the past two weeks, I averaged more than 150 interactions a day with Muse.

It independently planned two trips for me: a three-day weekend getaway and an extended journey of over a week. It handled the entire execution chain, from scouting flights and hotels to comparing prices and reviews, checking local event schedules, and sequencing itineraries around flight times and hotel checkouts. It even contacted event organizers to reserve slots and process payments. Along the entire chain, I never had to intervene in any operational detail, except to make one phone call to a local vendor that only accepted reservations by phone.

What really hooked me was its performance on long-horizon tasks. I run several persistent sentinel tasks in the background, including tracking critical inbox updates, watching portfolio movements, monitoring price drops on specific consumer goods, and waiting out ticket drops for concerts and national park permit lotteries. I used to do this with custom scrapers, but Muse acts like a round-the-clock personal concierge, dropping the barrier to entry to zero.

Much like ChatGPT, Muse is a notable product innovation rather than a fundamental technological breakthrough. At its core, it is essentially OpenClaw with vastly better interaction design. OpenClaw proved to the tech world earlier this year what an agent with full life permissions could do, priming mainstream awareness in the process. But requiring an entire weekend to configure API keys, gateways, and custom skills doomed it from ever reaching the mass market.

What Meta truly got right was executing three packaging feats: compressing setup from a whole weekend down to a single download, making integrations with everyday life and payment rails built-in out of the box, and turning security from a user burden into an architectural guarantee by provisioning an isolated cloud virtual machine for each user to keep all actions contained within a sandbox.

I particularly appreciate Muse's observability design. Most agents rely on dumping walls of cold terminal logs, leaving users with the uneasy suspicion that a rogue black box is at work. Muse turned Tool Use into a dedicated interactive window. Directly inside the chat thread, users can watch the agent open pages and enter search queries in its cloud sandbox browser, with the ability to take over the mouse and keyboard at any moment.

A PhD friend mentioned to me that although Muse was never built for academic research, watching the agent's full chain of thought and actions unfold live in the browser during deep research made all the difference. That visual clarity does far more to build trust between humans and autonomous agents than any terminal log ever could.

So while everyone obsesses over foundation models, compute scale, and harness frameworks, sharp product interaction design remains immensely valuable in its own right.

Meta's AI Super Gateway

Judging by the aggressive promotional push from Mark Zuckerberg and Alexandr Wang over the past two weeks, Muse has clearly become Meta's next high-stakes bet in the AI race. Whether it carries the hopes of the entire company remains an open question. Yet because none of Meta's existing social media products have shown any sign of becoming the super gateway for the AI era, Muse has undoubtedly given Meta a fighting chance in this contest.

Can Meta truly break out on the back of Muse? I believe the path ahead will be exceptionally grueling.

Muse's current growth is driven more by narrative momentum than overwhelming numbers. It topped the US App Store free chart on September 18, logging just over 730,000 total downloads across the US in its first ten days, or roughly 40,000 per day. That figure is telling. Compared with Threads pulling in 4.3 million downloads on day one, hitting number one on just 40,000 daily installs reflects less of a viral breakout and more of how low the mobile app market's baseline volume has fallen by the fall of 2026.

Still, Muse launched cold as a standalone brand, without aggressively funneling users from Instagram or Meta's other flagship apps. Furthermore, the macOS version released on September 17 was distributed directly through the web. Those desktop power users, who granted the agent direct access to their local files and inboxes, never showed up in App Store rankings.

Rankings aside, Muse confronts a much heavier cost structure and a complex business model paradox.

Because of its dedicated sandbox architecture, Meta must keep a virtual machine running continuously in the cloud for every active user. This stands in sharp contrast to its traditional asset-light model of feed distribution, which generated extraordinary gross margins. On top of that, Meta is burning through compute with aggressive token subsidies during this initial land-grab phase.

Under the hood, Muse runs on Meta's proprietary Spark 1.3 model, which lists on OpenRouter at $1.25 per million input tokens and $4.25 per million output tokens. Muse offers a free tier of 100 million tokens per week. Assuming a typical agent input-to-output ratio of roughly 9:1, that represents a subsidy of about $155 per user each week, which equates to nearly $8,000 a year in compute value. For heavy users like myself who interact over a hundred times a day and run continuous background tasks, consuming tens of millions or even 100 million tokens a week is effortless. The annual infrastructure and compute expense for a single user can easily run into thousands of dollars.

For now, capital markets are buying into the narrative. JPMorgan lifted its price target to $820, and Goldman Sachs projected that converting 10 million paying subscribers could yield $2.4 billion in annual revenue. Muse currently runs ad-free, with Meta testing a transaction take-rate model on purchases completed by the agent. If that model works, it will mark the first time Meta captures value outside advertising impressions. For a business where 97% of revenue still hinges on ads, this represents a structural transformation, which is why Wall Street is so energized. Yet if transaction commissions fail to materialize, Meta will likely retreat to its familiar advertising playbook. Relying on heavy token subsidies to sustain engagement simply cannot last for a company without a clear cost advantage in compute.

The Pivot from Intelligence to Trust

I tend to view Muse as a horse Meta simply had to field in the race for the next super gateway. It may not run the fastest or the farthest, but in a contest this consequential, Meta must have its own horse in the race.

Compared with the four incumbent AI gateways, namely ChatGPT, Gemini, Claude, and X, Muse targets a different space, leaning light on productivity, heavy on everyday use, and prioritizing trust over pure intelligence. The primary asset of an intelligence gateway is model capability, but performance leads tend to disappear every one to three months. By contrast, the defining asset of a trust gateway is how much access to private, everyday accounts users are willing to entrust to it. Foundation models are easily substituted, but once users deeply entrust their email, calendar, home address, and payment details to a specific node, switching costs rise dramatically.

Within this framework, how Muse measures up against the four major gateways becomes clear:

•ChatGPT: Its Agent Mode directly rivals Muse, but functions largely as an add-on within a massive subscription ecosystem. Its edge lies in its high intelligence ceiling, while its primary limitation is the absence of a native communication flow where it can permanently reside.

•Gemini: The permissions Muse works so hard to secure belong to Gemini by birthright. Google never needed third-party connectors, having already built Gemini into Gmail, Google Calendar, Google Maps, Android, and Google Pay. Google also commands deep defensive moats through Android and Chrome, and it has already rallied major retailers and payment networks into the Universal Commerce Protocol to prepare for agentic checkout. Its main vulnerability remains sluggish product execution.

•Claude: By positioning Cowork squarely around code, deep knowledge work, and enterprise document collaboration, it pursues an entirely different lane. For that reason, the emergence of Muse poses minimal threat to Claude.

•Grok: Although xAI's August release of Grok Bot also adopted a per-user cloud sandbox model, making it the closest philosophical match to Muse, X lacks depth in everyday consumer services. Consequently, winning user trust across sensitive tasks like managing finances, schedules, and daily affairs remains an uphill climb.

Within the same niche, Silicon Valley has also produced another high-profile startup with a product formula that closely mirrors Muse, called Instinct. Founded by a 23-year-old former Sierra engineer, it reached a multibillion-dollar valuation before even launching a public beta. Unlike Muse, Instinct trades radical permissions for capability, giving the agent a dedicated inbox, outbound calling power, and continuous screen and keystroke capture.

This differentiated landscape gives Meta an opening, yet Meta enjoys no natural high ground when it comes to trust. On distribution alone, its base of billions of daily active users across WhatsApp, Instagram, and Messenger easily outclasses pure AI startups. But when money, health, and private correspondence enter the picture, that distribution strength can quickly run into a wall of skepticism.

An Oppenheimer survey of 1,500 American consumers found that only 8% were willing to entrust passwords to Meta, compared with 30% for Google, 23% for Apple, and 16% for ChatGPT, while 58% refused to hand credentials over to any AI system whatsoever. A recent Visa study told a similar story, showing that just 23% of US consumers trust an AI agent to handle payments, while over 60% refuse to let an AI manage money without human oversight.

By relying on isolated virtual machines, mandatory browser visibility, and human confirmation for sensitive actions, Muse is essentially spending heavy compute resources to compensate for Meta's underlying trust deficit.

Where Is China's Muse?

I have not seen any consumer agent assistants like Muse or Instinct emerge in China. Curious about this absence, I consulted several AI agent founders across China's tech ecosystem. Those conversations convinced me that China not only lacks a direct counterpart to Muse today, but also that a consumer agent in China will almost certainly never take this form.

Muse can thrive in the West because everyday digital life runs on open, standardized protocols such as OAuth, IMAP, and CalDAV. Chinese consumers, by contrast, live inside the heavily fortified walled gardens of dominant super apps like WeChat, Alipay, Taobao, and Meituan. WeChat guards its private domain more fiercely than anyone in the industry, and it will never grant authorization tokens allowing third-party agents to act on behalf of users. Meanwhile, the tight grip that WeChat Pay and Alipay maintain over payment rails eliminates the room third-party agents need to survive on transaction take-rates.

This fundamentally different soil has forced China down an entirely opposite evolutionary path. In the United States, the race for personal agents is fought in the cloud and at the application layer. In China, that contest will unfold at the operating system level or inside the walled gardens of the super apps themselves.

In tracking Muse's staying power over the coming months, I would mainly focus on four indicators, namely true Day-30 retention, paid conversion rates, the viability of transaction-based commissions, and whether its Skill ecosystem opens up to third-party developers. Speculative valuations tend to feed on sweeping promises about managing an entire life. In reality, the most durable commercial returns come from solving bounded, recurring frictions with visible before-and-after states.