Ordinary users are “farming shrimp” with OpenClaw — and the tricks are surprisingly low-tech

Big tech jumped in. So did regular people.

In days, China’s major cloud and internet firms piled into the OpenClaw craze: Alibaba (阿里巴巴) rolled out CoPaw, HiClaw and JVCClaw; Tencent (腾讯) and ByteDance (字节跳动) launched QClaw and ArkClaw on the same day; Baidu (百度) and Huawei (华为) quickly followed. Cloud vendors began offering “one‑click” deployments to convert open‑source agent buzz into platform hooks. But the frenzy has not been without cost. It has been reported that social media is already showing negative cases — deleted files and system crashes — and the National Computer Network Emergency Response Technical Team has issued a public risk notice; several state firms and banks have restricted installing such agents on office machines.

What ordinary users are actually doing

A Huxiu report that spoke with three heavy OpenClaw users shows the reality is less sci‑fi and more pragmatic. An early‑stage investor, known as Xianghai, runs a hybrid setup: a cloud instance (KimiClaw) that scrapes financing news and a local agent that digests thousands of past BPs and research files to generate strategy reports and fast due‑diligence summaries. Minutes, not days, now separate a pitch deck from a first‑pass background check. Another user, entrepreneur Feigui, wires five machines (local laptops plus VPSes) to run daily morning briefs, paper‑summaries and a B2B industry radar — cutting a 50‑source workflow into roughly two hours. Costs? He estimates about ¥5,000 a year for model access and roughly ¥4,000 for three VPS nodes.

The limits: stability, controllability and UX

Those gains come with caveats. Users report “drift” in outputs, brittle long‑running sessions and poor mid‑task intervention — agents often execute end‑to‑end without easy user steering. Feigui’s idea of a “claw army” — specialized agents for planning, drafting and polishing that swap models by role — works in concept but stumbles when tasks grow non‑linear. Floatboat.ai founder Shaoqing has built a product to lower the deployment bar and argues the real bottleneck is interaction design: chat‑based control fragments workflows, forcing users to download, open and edit outside the agent’s context. In short: powerful automation, poor observability.

Productive tool or overblown hype?

So which is it — a nascent productivity revolution or an overhyped tech party? The answer may be both. Against a backdrop of export controls and a national push for AI self‑reliance, Chinese vendors are racing to package agent frameworks into mainstream tools. But governance and UX lag behind. It has been reported that regulators and large enterprises are already tightening controls. For now OpenClaw and its clones deliver clear productivity gains for technically curious users — but widespread, safe adoption will require better controls, clearer audit trails and interfaces that let humans meaningfully intervene.