The "China Muse" exists as a sourced report; it does not yet exist as a product. A Caixin-sourced report relayed by Phoenix Tech on Oct 11, citing people inside ByteDance, said the company accelerated work on a personal agent once Meta's Muse drew attention: the project, codenamed Spell, started in April, continues the Doubao phone assistant's personal-agent direction from 2025, and is run by the Doubao phone-assistant team together with the Doubao conversation team, with a plan to fold it into Doubao and ship it as soon as possible[1]. Aggregator reports add that the project was named Xiaodou this summer and is planned as a standalone app, still in internal testing[2].
The reports stop short of any official ByteDance confirmation, and the aggregator report notes that no official launch date has been announced[1][2]. That is the answer to the question circulating on Zhihu: a credibly reported project, no product.
The record already proves the other half of the story, and that half is usable now. Muse's first two weeks show what personal agents can and cannot do; the platform fight around Muse shows what will decide the category; and ByteDance's own office agent has been shipping since August. If you track one thing, track the enterprise piece you can already test; the consumer app, for now, exists only as a report.
Muse's first two weeks: real traction, visible limits
Muse went live on Sept 8. It is built to execute tasks rather than answer questions: triage and send email, make calls, book travel, fill in web forms, handle bills, buy things, and keep working in the background after the user closes the app[3]. It runs on iOS, Android, the muse.ai website and WhatsApp[4].
The pricing, as launched, in US dollars: a free tier with about 100 million tokens a week plus a cloud virtual machine; Power at $20 a month with 500 million tokens a week; Maximum at $100 a month with 3 billion tokens a week[3]. The app shows remaining quota and prompts an upgrade once the free allowance runs out. Zuckerberg has framed monetization around volume rather than subscription revenue: Meta wants Muse free because the product "helps you make money and save money — that's how it pays for itself" (translated), with a small commission on transactions and paid premium tiers on top[3]. For anyone costing this out, the components are the subscription price, the token quota that caps how much work each tier can do, and any transaction commissions Meta may charge.
Traction is real, with mixed sourcing. By Sept 21, under two weeks after launch, Muse topped both the US App Store and Google Play free charts[3]; IT Home's report puts the App Store topping within the first week[4]. App-analysis firm Apptopia estimates about 2.8 million installs in the first 12 days, 1.8 million of them in North America against ChatGPT's 1.3 million at the same stage after its mobile launch. Third-party estimates put US mobile daily active users at roughly 642,000 against ChatGPT's 231,000, with 359,000 on iOS alone[3]. All of it is tracker data, not Meta figures.
The limits are as well documented as the growth. Business Insider's Katie Notopoulos spent a week handing Muse real tasks with her mailbox, bank card and Apple Health data attached. Simple, well-defined work went fine: filtering important mail out of press releases, daily summaries, analyzing subscription spend. Multi-step work broke repeatedly: Muse could not get past a Google-account login for a school form routed through ParentSquare, a Girl Scouts membership renewal timed out at the final payment step, an annual-checkup booking stalled on insurance information until the agent suggested calling the clinic, and a Walmart order where she asked Muse to stick to free-shipping items and it missed the extra same-day delivery fee, so she canceled the task[3]. Meta spokesperson Jimmy Raimo said some of those stops are deliberate: when charges look unusual or a login gets rough, Muse pauses and asks the user rather than pushing through[3].
Amazon said no, Shopify said yes: where the category gets decided
The dispute shaping this category more than any benchmark is about permission: who consents, who discloses itself, who holds the credentials. Meta's launch materials state the mechanism. Where a service offers a public API, Muse connects with the user's credentials; where there is no API, it "can use the service like a human, through a browser" (translated)[4].
Amazon blocked Muse from shopping on its site after trying and failing to persuade Meta to exclude Amazon on its own. Amazon's stated reasons: the access was never licensed, Meta gave no advance notice, Muse does not identify itself while browsing, and it appears to collect and store account credentials. From the night before the Sept 21 report went out, users saw a banner citing Amazon's terms of use, and Amazon has asked Meta to remove Amazon from Muse's service list[4]. Meta's stated position runs the other way: Muse cannot read passwords or payment information, credentials sit in separate secure storage that Muse can invoke without seeing, sensitive operations require another confirmation, and a separate monitoring agent reviews everything Muse sends out[3][4].
Two things make this more than a scuffle. One is money: Amazon's advertising revenue topped $68 billion last year, and that business depends on users browsing and seeing sponsored products — the behavior an agent replaces with a task[4]. The other is the legal frame. Amazon sued Perplexity over its Comet browser and won a temporary injunction in March, but on Aug 4 the Ninth Circuit ruled that under federal anti-hacking law the party accessing Amazon's servers is the end user, not the AI company, which voided the injunction; a rehearing was denied on Sept 10. What platforms are left with is contract and terms of service — the exact instrument that banner invokes. Amazon has also blocked shopping agents from Google and OpenAI, and its own Buy for Me feature — by Amazon's own description — identifies itself on external sites and lets brands opt out: the disclosure standard it wants Meta to meet[4].
Shopify made the opposite choice the same week. On Sept 21 it put Shop Pay into Muse so the agent can check out across Shopify stores, and CEO Tobias Lütke described it as deep collaboration. Deutsche Bank's read is that the deal makes Shopify's catalog, merchant connections and Shop Pay the infrastructure agents need, rather than getting bypassed by them, while flagging the unsettled questions of how agents will charge and whether they take commissions[3].
The economics are the second gate. An analysis cited in Jiemian's round-up of the space sets out Muse's revenue problem: either compute costs fall substantially, or the average user has to buy more than $1,000 a month through the agent, well above current e-commerce behavior[5]. Microsoft's Sept 25 move runs the other way: it merged consumer and enterprise Copilot into one product and pivoted to enterprise customers, which Bloomberg read as a retreat from the consumer personal-assistant race[5].
ByteDance's verifiable position: Doubao Work, not Xiaodou
What ByteDance has actually shipped is an office agent. Doubao Work launched Aug 25 for productivity scenarios: it decomposes tasks, calls tools and runs workflows end to end across proposals, meeting notes, data analysis and document work; with user authorization it operates the local PC and browser, and long tasks keep running in the cloud even if the machine shuts down[6]. It is deeply integrated with Feishu and reuses Feishu's permission and isolation mechanisms — the enterprise-grade part of this story, and the reason ByteDance's position looks different from a consumer agent racing across open websites. It grew out of Doubao's office-task mode, works through the Doubao PC client, and new users get a 30-day free subscription; ByteDance has not published the price after the trial[6].
The organizational moves preceded it. In late July ByteDance folded the Feishu product team into Doubao, then pulled TRAE and Coze into the same system, and China Business Journal's report frames the competition as having shifted from model capability to real workflow implementation, with Alibaba and Tencent fielding office agents of their own[6].
The consumer track remains the reported part. Spell's lineage goes back to a December 2025 technical preview of the Doubao phone assistant running on a ZTE-Nubia engineering device, nine months before Muse launched — which fits "accelerated" better than "copycat"[2]. The Caixin-sourced report says the plan is to combine the agent with Doubao and ship fast[1]; per the aggregators, Xiaodou is the name, a standalone app is the form, and internal testing is the state[2]. Nothing beyond that is confirmed.
Tencent's move is documented from two directions. The Caixin-sourced report has Tencent launching LightVela at the end of September as a "cloud personal assistant" reachable through WeChat, still in a developer-experience stage[1]. LightVela's own documentation describes a 7×24 cloud agent on Tencent's Lighthouse platform running a Hermes Agent instance, reachable over WeChat, QQ, WeCom, Feishu and DingTalk, configurable across model providers, and sold as subscription plans in two series with monthly or quarterly billing, model-credit packs, and real-name authentication before the first purchase — the vendor's own description, and the documentation lists no amounts[7].
Until ByteDance confirms the project or the standalone app appears, "China Muse" stays a report: well sourced, with a named team and a start date, and still no official ByteDance confirmation in any of the reports[1][2]. Watch which platforms let agents act and which block them, rather than the model demos — Amazon's ban, identity-disclosure demand and terms-of-use banner on one side, Shopify's agent-checkout deal on the other; that choice will decide how much of the personal-agent promise survives outside walled gardens. And test what already ships: Doubao Work, reached through the Doubao PC client with its deep Feishu integration, is the piece an enterprise reader can try now — 30 days free for new users — reusing Feishu's permission model and, under user authorization, browser and local-machine actions. That is where ByteDance's advantage is provable today, and where the "personal agent as new direction" question gets answered first.