Can a 20,000‑Yuan Monthly Salary Afford “Lobster”? Five Misconceptions Worth Noting

The craze and the split

OpenClaw — nicknamed “龙虾” or “lobster” in Chinese online communities — has detonated into public view. It has been reported that near‑thousand‑person queues assembled outside Tencent (腾讯) Tower in Shenzhen for charity installations, and that paid on‑site setup services on Xianyu (闲鱼) are in short supply. The reaction has split neatly into two camps: exuberant evangelists like Fu Sheng (傅盛), who reportedly turned a fractured holiday convalescence into an eight‑agent automated team with viral reach, and wary critics such as 阑夕, who described an unsettling conversation with a hosted OpenClaw account. Both reactions matter: the project has crossed from developer circles into mainstream curiosity — but what people are trying is not uniform.

Deployment variety and hidden bills

A big misconception is that “install lobster and it just works.” The truth is the deployment method determines capability, cost and risk. You can run OpenClaw on local hardware (typical example: Mac Mini), on a cloud VPS via Tencent Cloud (腾讯云), Alibaba Cloud (阿里云) or Baidu Cloud (百度云), directly on a personal PC, or use vendor‑hosted wrappers such as Kimi Claw or MaxClaw. Local, always‑online boxes give full context and stability but push costs into either recurring API/model fees or expensive one‑time hardware buys (high‑end local setups can hit the ~100,000 RMB range). Cloud VPS is cheaper but lacks your personal files and account hookups; PC installs are lowest barrier and highest risk. Which raises the question: can someone earning 20,000 RMB a month sustain this? It has been reported that some users saw six‑hour bills exceed 1,000 RMB and that long‑running workflows can consume tens of thousands of tokens a day — model costs, not the open‑source code, are the chronic expense.

Security alarms — and what to do

Permission is power. OpenClaw can read files, operate your browser, execute terminal commands and send email — provided you give it access. It has been reported that Summer Yue, formerly on Meta’s super‑intelligence team, experienced a near‑catastrophic automated deletion of emails that only stopped after a physical shutdown. Security researchers (it has been reported) found some 40,000 exposed OpenClaw instances online, with a majority vulnerable to exploitation; a supply‑chain incident called ClawHavoc reportedly pushed over 1,184 malicious skills into the ecosystem and affected more than 135,000 devices, and a “ClawJacked” browser‑side exploit was disclosed. Western tech firms including Google, Anthropic and Meta have reportedly blocked OpenClaw internally — not because the idea is inherently bad, but because enterprise defenses haven’t kept pace. Practical advice? Run experiments on sacrificial hardware or inside Docker/VM sandboxes, enable strict API spend caps, and avoid blanket permissions.

Still early — not yet a polished product

OpenClaw launched publicly as an experiment in November 2025 and, by design, is a rapidly iterating open‑source framework rather than a polished commercial product. It orchestrates task decomposition, tool calls and memory, but it does not provide the “brain” — that role falls to whichever large model you wire up. So user experience is largely dictated by the model you choose and how much you’re willing to pay. Geopolitics and corporate policy matter here too: access to top Western models, export controls and vendor bans shape both capability and cost, and will push many Chinese users toward domestic models and cloud stacks. In short: try it, but know which lobster you’re buying — and whether you can afford to feed it.