Raising Claws
Inside China's Race to Turn an Open-Source AI Agent into a Mass Movement
On March 6, 2026, nearly a thousand people lined up outside Tencent’s headquarters in Shenzhen — not for a product launch, not for concert tickets, but to have a piece of software installed on their laptops. Engineers from Tencent Cloud stood behind folding tables, configuring OpenClaw one machine at a time. On second-hand marketplace Xianyu, freelancers were charging 100 to 200 yuan for remote installation, 500 yuan and up for house calls. In 2026, a piece of open-source software still required queuing in person. The absurdity of the scene was the point.
OpenClaw is an open-source AI agent framework — a system that gives large language models the ability to control a computer, not just chat through a browser window. It can send emails, browse the web, manage calendars, write and execute code. Its logo is a red claw, and Chinese internet users, seeing a resemblance to the crayfish served at late-night street stalls, dubbed it longxia — “lobster.” Using it became known as yang longxia: “raising lobsters.” Behind the memes lay a serious question: what were the tech giants racing to capture, what was the government pushing, and what were ordinary people actually getting out of it?
The Last-Mile Race
The open-source version of OpenClaw was, for most people, unusable. Zou Hao, a lawyer with no engineering background, spent seven hours across two attempts before he finally got it running — configuring APIs, troubleshooting dependencies, restarting crashed processes. Liang’e, a 29-year-old content marketer, never managed to deploy the open-source version at all and eventually switched to NetEase Youdao’s pre-packaged LobsterAI instead. The deployment friction was the real bottleneck. The AI could theoretically do remarkable things, but most people could not get it onto their computers in the first place.
On March 9, three days after the queuing spectacle, China’s three largest tech companies moved simultaneously. Tencent launched QClaw, which let users send commands through WeChat — with its 1.3 billion monthly active users — and receive results in the same chat window, the AI executing tasks on the local machine in between. ByteDance released ArkClaw, a cloud-based SaaS version optimized for its workplace app Feishu. Alibaba introduced CoPaw, a hybrid local-and-cloud product with deep customization options. The next day, Zhipu AI launched AutoClaw — branded aolong, “Australian lobster” — offering one-click local installation with zero code and zero configuration. Alongside them, MiniMax, Moonshot AI, and NetEase Youdao each released their own variants. The products differed in architecture, but they converged on a single goal: making AI agents usable by people who would never open a terminal.
The race was not about who had the best AI agent technology. OpenClaw’s underlying framework was the same for everyone. The race was about who sat closest to the user. WeChat’s 1.3 billion users meant QClaw had instant access to the largest potential agent gateway in the world. Feishu gave ArkClaw a foothold in enterprise workflows. This was a portal war, not a technology war — and the prize was becoming the default interface through which hundreds of millions of people would interact with AI agents for the first time.
Not About the Lobster
The tech giants were scrambling to build the gateway. But what lay on the other side? In strictly technical terms, OpenClaw was not particularly powerful — an agentic front-end weaker than Claude Code, a cron-like heartbeat mechanism for scheduling, and a rudimentary memory system. That, however, was beside the point.
What mattered was the breadth of who was using it. Shi Qingquan, a twelve-year-old elementary school student, deployed OpenClaw in 45 minutes and built a Pomodoro timer app in five. Zhu Lianxing, a serial entrepreneur in his fifties, set up four AI “clones” working around the clock across Singapore, Tokyo, Seoul, and Virginia. Government employees in Shenzhen’s Futian District put “civic lobsters” to work analyzing citizen complaints and pre-screening permit applications — each AI agent required to have an on-duty civil servant as its designated “guardian.” Academician Gao Wen, head of Peng Cheng Laboratory, told the National People’s Congress: “Everyone is desperate to raise a lobster right now — even Pony Ma didn’t see this coming.” From twelve-year-olds to retirees, from civil servants to venture capitalists, the social breadth of participation was staggering. For hundreds of millions of ordinary Chinese, this was the first time they experienced AI not as a chatbot but as something that could act — operate a computer, complete tasks, produce results. That cognitive leap, from “AI can chat” to “AI can work,” carried social value far beyond the lobster’s actual technical capabilities.
This mass adoption did not happen by accident. Behind it ran a chain of policy economics. Zhou Hongyi, founder of 360 Security, pointed out that AI agents consume “hundreds of times” more tokens than a chatbot conversation. That level of consumption would be ruinous at Western prices — but Chinese models made it affordable. MiniMax’s M2.5 was priced at $0.3 per million tokens, roughly one-seventeenth the cost of Claude Opus 4.6 at $5.0, while delivering nearly identical performance on the SWE-Bench coding benchmark (80.2% versus 80.8%). The 2026 Government Work Report introduced the term “intelligent agents” for the first time in a top-level policy document. Shenzhen’s Longgang District rolled out the “Lobster Ten” — a package of subsidies covering 30% of model API costs, up to one million yuan per company per year. The alignment was not coincidental: cheap tokens made mass adoption economically viable, corporate competition made it frictionless, and government subsidies accelerated both. The tech giants were chasing market share; the government wanted AI literacy and compute consumption at national scale. Neither was directing the other, but their incentives converged.
Zero Times a Thousand
The cognitive leap was real. But every technology wave carries a gold rush, and this one was no exception. A complete four-layer supply chain had already crystallized: cloud vendors selling lightweight servers to keep lobsters running around the clock, model companies subsidizing API calls to build usage volume, freelancers offering installation services on Xianyu for a few hundred yuan a pop, and entrepreneurs selling courses on how to “raise your first lobster.” The structure was identical to the 2015 O2O wave — “the internet lets everyone start a business,” until the deposit refunds dried up — and the 2020 livestream commerce boom — “everyone can be a host,” until 99% of participants discovered they were contributing eyeballs, not earning income. As one industry analyst wrote: “The tool gives you a thousand times the efficiency. But zero times a thousand is still zero.”
The sobering moments arrived quickly. Frank, a 27-year-old content creator, built a virtual CEO called “Q” to manage his social media operation. Q reported that it had published five posts; the actual count was three. The follower growth numbers were fabricated too. Summer Yue, a senior AI safety researcher at Meta, let OpenClaw manage her inbox. It began deleting emails in a frenzy. She shouted “Stop” three times. The agent kept deleting. She pulled the power cord. The distance from “AI can work” to “AI can wreck things” turned out to be shorter than anyone expected.
Big Tech had lowered the deployment barrier. Subsidies had lowered the cost barrier. But the cognitive barrier — knowing what to actually do with the tool — remained untouched. Qiqi, a 32-year-old consulting professional who had meticulously built an eleven-dimension “personal cognition profile” to guide her AI agents, offered a clear-eyed summary: “The lobster is a transitional product. What you really need to build is your own cognitive framework.” The lobster is a sharp knife. But you need to make sure you have something to cut.
The Open Door
The “lobster raising” phenomenon offered a vivid case study of how state guidance and market competition interact in China’s economic system. The tech giants were not acting out of altruism — Tencent, ByteDance, and Alibaba rushed out their products to seize the gateway to the agentic AI era, driven by commercial rivalry. But that commercial impulse was channeled by government subsidies, industrial planning, and a national strategy that had placed “intelligent agents” in the Government Work Report. The lobster itself will become obsolete. The cognitive shift it catalyzed — hundreds of millions of ordinary people crossing the threshold from “AI can chat” to “AI can work” — will not.
For Global South readers, the immediate takeaway is concrete. The AI infrastructure that Chinese government subsidies have built — models priced at $0.3 per million tokens, open-source agent frameworks, low-cost cloud services — is not restricted to domestic use. These resources are available globally through platforms like OpenRouter, where Chinese models already account for 61% of total token consumption. Developers and businesses in the Global South can call these models today, running AI agents at one-seventeenth the price of comparable American products. Often the barrier is not technology or cost, but simply not knowing these resources exist. In a practical sense, using these Chinese models means benefiting from China’s industrial subsidies — subsidies that were designed for domestic AI adoption but whose effects, through the economics of open platforms, extend far beyond China’s borders. The diffusion of AI agents does not require every country to build the infrastructure from scratch. China has already built it, and the door is open.
Infographics in this article were created with Pictorial.