OpenClaw in China: A Practical Guide for Outsiders

Here’s a take you won’t hear at a Beijing meet-up: the OpenClaw craze in China has almost nothing to do with OpenClaw itself. Strip away the lobster plushies and the queues at Tencent HQ, and what’s left is a much older story – a cheap API stack, a cottage industry of installers, and a government that decided the technology was useful before it decided whether it was safe. If you’re approaching OpenClaw China as an outsider, copying the Western tutorial won’t get you the same tool that Chinese users are running.

This guide is for the practical case: you want to understand what’s actually different about the Chinese deployment, set one up yourself using that stack, and avoid the specific traps that triggered the wave of mass uninstalls in March 2026.

What “OpenClaw China” Actually Means

OpenClaw itself is the same software everywhere. According to Wikipedia, OpenClaw bots run locally and are designed to integrate with an external large language model such as Claude, DeepSeek, or one of OpenAI’s GPT models, with functionality accessed via a chatbot within a messaging service like Signal, Telegram, Discord, or WhatsApp. It was created by Austrian programmer Peter Steinberger and released in November 2025.

The Chinese version of the stack diverges in three concrete ways: the model, the messaging layer, and the deployment surface. Chinese developers adapted OpenClaw to work with the DeepSeek model and domestic messaging super apps such as WeChat. The cloud layer is the bigger story – Alibaba Cloud, Tencent Cloud, ByteDance’s Volcano Engine, JD.com, and Baidu have all embraced OpenClaw or a spinoff, each offering one-click deployment that sidesteps the GitHub-and-CLI ritual that Western users go through.

That last point matters. The Diplomat reported that Chinese official media, in explaining why no equivalent OpenClaw craze emerged in the West, pointed to strict data privacy regulations and higher computing costs, framing cheap APIs as a comparative advantage unique to China. Read that as the mission statement of the entire Chinese OpenClaw economy.

The Stack You’ll Actually Use

If you’re configuring OpenClaw to mirror what’s running in China – for benchmarking, research, or just to use a cheaper model backend – here’s the rough shape of it.

The spinoffs aren’t separate products in any meaningful sense – most are skinned OpenClaw plus a domestic model and a WeChat connector. KimiClaw reached 1.6184 million users in its first month, which gives you a sense of how fast the market splintered into branded variants. One data point that doesn’t get quoted often: China’s daily token usage climbed from 100 trillion at the end of 2025 to 140 trillion by March 2026, according to the China National Data Administration via Wire China. That jump is driven by agent loops calling APIs continuously, not by people typing into chat windows.

Setting It Up: The Honest Version

The fastest path is the official installer. From the official OpenClaw docs:

# Bun / npm install
bun add -g openclaw@latest
openclaw onboard --install-daemon

# Verify
openclaw --version
openclaw doctor
openclaw gateway status

The onboarding wizard asks for an API key. This is where outsiders diverge from the Chinese setup. Instead of pasting an Anthropic key, point it at a DeepSeek-compatible endpoint. Per Ollama’s deployment guide, agents work best with at least 64k context length – DeepSeek’s V3 chat models clear that bar comfortably and cost a fraction of Claude per million tokens.

If you want the Chinese cloud experience without being in China, the simulation is straightforward: spin up any Linux VPS, install OpenClaw via the official script, and route requests through a DeepSeek API key. You won’t get the WeChat integration that QClaw provides, but you’ll get the cost profile that drove the craze.

Pro tip: Before you connect any messaging channel, run openclaw doctor and confirm the gateway is bound to 127.0.0.1, not 0.0.0.0. The setup guides that go viral on Douyin tend to skip this – and an exposed gateway is exactly the configuration that Chinese cybersecurity agencies flagged as a remote-takeover vector.

The Pitfalls Tutorials Skip

The China story is also a story of what goes wrong. Three specific traps are worth your attention.

1. The background bill. OpenClaw isn’t a chatbot you open and close. It runs around the clock, calls model APIs hundreds of times per day, accesses local files, and operates browsers. Wire China reported that users racked up hefty bills after keeping their AI agents running in the background on their computers. This isn’t a hidden pricing trap – it’s a direct consequence of agent loops calling out to a paid API while you sleep. Set a hard spending cap in your provider’s dashboard before you connect the key.

2. Unvetted third-party skills. OpenClaw’s skill repo is the real attack surface. Skills ship as directories with a SKILL.md file – they can read your shell, access files, and call external URLs. The public repo has no meaningful vetting process. Cisco’s AI security research team tested a third-party OpenClaw skill and found it performed data exfiltration and prompt injection without user awareness. Treat installing a community skill the way you’d treat running a random shell script that asks for your home directory.

3. Destructive actions without confirmation. The most-quoted Chinese horror story is the Shanghai consultant whose QClaw deployment permanently erased dozens of documents, including reports he had written for his clients, when he asked it to organize files. This isn’t a QClaw-specific bug – it’s the default agent behavior. Scope filesystem access to a dedicated project directory, not your home folder, and set a hard spending cap in your API provider dashboard before the daemon starts running. Both settings live in ~/.openclaw/openclaw.json.

One of the project’s own maintainers said it more bluntly than any journalist. A maintainer known as Shadow warned on Discord that “if you can’t understand how to run a command line, this is far too dangerous of a project for you to use safely.” Whether or not that’s true, it’s at odds with Baidu’s installation queues for retirees.

OpenClaw vs. the Domestic Alternatives

Which version should you actually run? For most non-Chinese users: vanilla OpenClaw with a DeepSeek key. QClaw, KimiClaw, and MaxClaw are useful only if you live inside their messaging platforms – they’re not technically superior, just better-integrated with WeChat or Lark. Tencent’s WorkBuddy, MiniMax’s MaxClaw, and Moonshot’s KimiClaw are the prominent branded variants, but their differentiation is mostly distribution.

The deeper question: is the agentic loop worth the complexity at all? The kind of automation OpenClaw offers has spurred hopes of efficiency gains in China, where 93% of respondents to a 2025 KPMG survey (via CNN) said they already use AI in their work. The US has adopted a warier stance – only 35% said the benefits outweighed the risks. Two countries running roughly the same software are reaching opposite conclusions about whether to trust it. Neither side has the better answer yet.

FAQ

Is OpenClaw legal to use in China?

For individuals, yes. In March 2026, the Chinese government moved to restrict state agencies, state-owned enterprises, and banks from using OpenClaw, citing security concerns such as unauthorised data deletion and leaks, and excessive energy usage. Private use is unrestricted, which is why the consumer craze continued even after the government warnings.

Why are Chinese users running OpenClaw so cheaply compared to Western users?

Token costs, mainly. The Chinese stack calls DeepSeek or Kimi instead of Anthropic or OpenAI – an order of magnitude cheaper per million tokens. On top of that, local government subsidies pile on: Shenzhen’s Longgang district offered grants of up to 10 million yuan ($1.4 million) for “one-person companies,” with similar programs in Wuxi targeting robotics use cases. A Chinese solo operator can run an always-on agent loop for what a Western user pays for a hobby ChatGPT subscription.

Did Peter Steinberger really sell OpenClaw to OpenAI?

Not quite. On February 14, 2026, Steinberger announced he would be joining OpenAI, and that a non-profit foundation would be established to provide future stewardship to the project. The software remains open source – Steinberger moved, the project didn’t.

Next step: grab a DeepSeek API key, set a $5 monthly hard cap on it, then run openclaw onboard --install-daemon on a throwaway VM (not your main laptop). Point the daemon at DeepSeek, ask it to summarize three RSS feeds into a single text file, and watch openclaw daemon logs while it runs. That’s the entire Chinese OpenClaw experience in one afternoon – minus the lobster plushie.