China's Tech Giants Race to Build AI's App Store, but Four Barriers Remain
Tencent, Alibaba, and ByteDance have each launched "Skill stores" for AI agents within months of each other, betting that controlling Skill distribution today will translate into ecosystem lock-in tomorrow — yet only one platform has attempted monetization, and four structural barriers are keeping the market firmly in demonstration mode.
The scramble began in March 2026, when Tencent, Alibaba, and ByteDance each rolled out dedicated Skill marketplaces on their respective agent platforms within weeks of one another. By May, Zhipu AI, Meituan, and Xiaohongshu had followed. The speed of entry signals industry-wide recognition that whoever controls Skill distribution in the AI-agent era could replicate the gravitational pull Apple's App Store exerted over the mobile internet decade — a platform that generated far more value through ecosystem consumption than through direct download fees.
Market feedback, however, is delivering an early reality check. With the sole exception of ByteDance's Coze platform, every Skill store currently operates on a fully free-distribution model. The only documented commercial transaction volume exists on secondary marketplaces such as Xianyu, where arbitrageurs repackage open-source Skills and resell them through information asymmetry — hardly the infrastructure of a thriving developer economy.
Three Distinct Playbooks Reveal Diverging Strategic Motivations
The competitive field breaks into three structurally different cohorts, each using Skill stores as a means to a different end.
Internet conglomerates are playing an ecosystem-retention game. Alibaba has embedded a Skill marketplace called "Xiaobaobao" inside its JVS Claw Agent assistant. Skills are free, but every invocation consumes compute credits billed against Alibaba Cloud — converting Skill adoption directly into cloud revenue. Tencent's SkillHub functions as a localized mirror of the overseas ClawHub developer community, but the company's true leverage lies in its WeChat Mini Program ecosystem, which already hosts millions of standardized service workflows. If Tencent successfully encapsulates those workflows as Skills, the monetization model mirrors Mini Programs: transaction commissions and advertising. Meituan, meanwhile, launched xia345 in April 2026 as an AI agent navigation portal — cataloguing over 20 agents and 7,000-plus Skills — followed in May by the public beta of its AI community platform Miyou, which now hosts more than 3,000 agents and over 40,000 Skills. Neither platform charges; the strategic calculus is that extended dwell time within Meituan's ecosystem generates incremental conversion for its core food delivery and in-store businesses.
Foundation model companies are using Skills as user-retention mechanisms to drive model inference volume. Zhipu launched AgentMore Skills Plaza on its Auto Claw platform in April 2026, integrating curated official picks, a Skill Hub, and open-source community modules with one-click, zero-token installation. Moonshot AI moved earlier, debuting Kimi Claw in February 2026 with a browser-based Skill library that allows direct deployment of Open Claw agents. The logic is structurally sound: foundation models are the runtime substrate for every Skill execution, so driving Skill adoption mechanically increases model call volume. As one agent engineer at a major model company told local media, "Skill is the bait; model invocations are the fish."
Content platforms are treating Skills as a new content category to monetize through traffic and advertising. Xiaohongshu's Red Skill, currently in closed beta, allows creators to attach Skill links directly beneath posts; viewers can copy installation instructions with a single tap. This inverts the conventional Skill discovery flow — from search-and-configure to browse-and-recommend — and plays directly to Xiaohongshu's algorithmic distribution strengths. The platform has no intention of taking a cut of Skill transactions; it is monetizing the attention that Skill-related content generates.
Four Structural Barriers Are Blocking Commercialization
The strategic rationale for each player is coherent, but the path to a functioning marketplace faces compounding obstacles that none of the incumbents has yet resolved.
Pricing without determinism is nearly impossible. The App Store succeeded because any user running the same application gets an identical experience. Skills lack this property: swap the underlying model, alter the context window, or run the same Skill on a different agent platform, and outputs can vary substantially. Without reproducible, measurable outcomes, no credible rating system can emerge — and without ratings, users have no rational basis for paying.
Opaque token costs distort total cost of ownership. Two Skills performing the same summarization task on the same platform can consume token volumes that differ by multiples, yet this information is invisible to users pre-installation. A user who pays for a Skill still faces an unquantifiable ongoing inference cost — a two-layer pricing opacity that suppresses willingness to pay.
Security incidents are raising the compliance burden. Multiple "Skill poisoning" cases emerged in the first half of 2026, where malicious Skills mimicked popular titles to exfiltrate user data. Platforms have responded with tighter review mechanisms, but the side effect is a higher upload barrier for legitimate developers. One independent developer reported that Xiaohongshu's review process restricts uploads to Markdown and TSD files, forcing complex Skills to be downgraded to basic prompts — a friction point that limits the quality ceiling of the marketplace.
The absence of standardized protocols prevents cross-platform portability. Inconsistent task-description conventions across developers generate model interpretation errors; inconsistent permission boundaries mean a Skill built for one platform rarely deploys cleanly on another. The "build once, distribute everywhere" vision that underpinned the App Store's developer appeal remains structurally out of reach.
Developer Economics Expose the Distribution Power Imbalance
Early monetization data from Coze offers the most concrete signal available. One independent developer reported that a paid Skill attracted six purchases on its launch day after receiving homepage placement — validating that latent demand exists. But homepage visibility is entirely platform-controlled and algorithmically opaque; the developer quickly lost front-page exposure and had no mechanism to purchase promotional placement. The asymmetry is stark: platforms capture the distribution chokepoint while developers bear the creation cost.
This mirrors the early App Store dynamic, but with a critical difference. In 2008, Apple's review and ranking systems, however imperfect, provided developers with at least a legible set of rules. Current Skill stores offer neither transparent ranking criteria nor paid discovery tools — leaving developer economics dependent on platform discretion.
Assessing the Gap to "App Store" Status
The App Store analogy is instructive precisely because it highlights what is missing. Apple's platform succeeded by solving for certainty: a defined SDK, a consistent runtime, a standardized review process, and a star-rating system grounded in reproducible user experience. Skills are, by design, personalized workflow automata — they resist the standardization that commercial distribution requires.
Industry observers identify two categories where monetization is most structurally viable in the near term: enterprise workflow automation (contract review, data report generation) where outcomes are measurable and corporate budgets are available; and high-intent consumer tools (resume optimization, application essay drafting) where users have clear willingness to pay for a specific output. Both categories share a common feature: the task has a definable success criterion, which partially mitigates the evaluation problem.
The competitive positioning of each cohort carries inherent limitations. Internet conglomerates are closest to high-value use cases but are unlikely to commit core engineering resources to what is, for them, a secondary feature. Foundation model companies have the deepest technical alignment but lack the ecosystem scale to attract a critical mass of developers. Content platforms possess the most powerful organic distribution channel — in the absence of standardized evaluation, peer recommendation and live demonstration are the dominant discovery mechanisms — but are furthest from the technical infrastructure needed to enforce quality and security standards.
Until the industry resolves the four structural barriers — output determinism, cost transparency, security governance, and protocol standardization — Skill stores will remain, in the words of one developer, "a display shelf where goods are on show but no one knows what to buy."
Related Coverage:
Alibaba's Qwen3.7-Plus Tops GPT-5.4 in Screen Tasks, Builds Apps Autonomously
Tencent's WeChat AI Agent Could Mark a Turning Point in Its AI Strategy
Meituan Loosens Its Grip on the Front End to Survive China’s AI Shift