"Lobster" Carnival AI "Shrimp Farming" Technology Guide: How to Raise Them and What Risks to Be Aware Of

As AI agent technology matures, the cyber “shrimp farming” trend is rising.

Here, “shrimp” refers to OpenClaw, an open-source AI agent framework named for its lobster-like icon. It can turn idle computers into tireless “digital employees” that automatically perform tasks.

Hu Yanping, a distinguished professor at Shanghai University of Finance and Economics and a scholar in intelligent technology industry, told The Paper that OpenClaw essentially takes over devices and humans, and can no longer be called an agent AI. Instead, it functions as a task-oriented operating system, which is why some believe OpenClaw is close to the next-generation OS prototype.

Currently, major domestic companies are actively developing the OpenClaw ecosystem. Various tech seminars remain popular, and installation services are in high demand. However, while enjoying efficiency gains, potential security risks and technical barriers should not be overlooked. Based on interviews with AI enthusiasts, experienced practitioners, and industry experts, The Paper has compiled an AI “shrimp farming” technical guide.

Hardware and Deployment: Cloud is easy to start, local has barriers

The first step in “shrimp farming” is not software installation but preparing the hardware environment. Xiao Cui, who recently installed OpenClaw on his Mac and Windows computers, told The Paper that any computer meeting the official requirements can try, whether it’s Apple or Windows.

Xiao Cui, an “expert lobster installer,” said that although the official instructions promote one-click installation, the process involves environment checks and dependency configurations that require technical skills. During installation, the system checks for missing components, and errors may occur depending on the user’s computer. If errors happen, users can consult GitHub discussions or official documentation for solutions.

Hu Yanping noted that while Windows installation involves more steps and details, the difficulty is not significantly higher than on Mac.

AI practitioner and blogger Lynn, who has been “shrimp farming” for over a month, told reporters that there are two deployment options: cloud and local. Tencent Cloud, Alibaba, and Volcano provide cloud servers where users can “farm shrimps” in the cloud. They can then connect the cloud-based “lobster” to apps like Feishu to automate tasks via chat. For local deployment, users need hardware like a Mac mini to run OpenClaw, using it as a local assistant.

One user built a complete weekly meal plan system covering all of 2026 (365 days), automatically updating shopping lists categorized by store and shelf, weather forecasts, and more.

Lynn pointed out that running “lobster” continuously on local hardware generates data stored on the computer, which requires sufficient memory and hardware resources.

As a result, the original Mac mini priced around 3,000-4,000 yuan sold out in many places, and second-hand rental services emerged.

Lynn uses both local computers and cloud services. She believes cloud deployment is simple and cost-effective for ordinary users, while local deployment offers better data privacy but requires some programming knowledge, making it less user-friendly.

Core Configuration: Model determines intelligence, skill workflows build capabilities

OpenClaw’s capabilities depend directly on the large model integrated. Hu Yanping believes that, in terms of device operation, agent support, and action understanding, the recently released GPT-5.4 is currently the most promising choice. For domestic users, three main Chinese models offer a good balance of accessibility and cost: Qwen-3.5, Kimi-2.5, and MiniMax M2.5. The first two perform better in multimodal tasks.

With a “brain,” it also needs “hands and feet”—OpenClaw’s official site offers over 10,000 Skills that users can choose to install.

Lynn mainly uses the “information scraping” Skill, especially “Twitter info scraping,” which she finds cost-effective for capturing trending news compared to official APIs. For complex tasks, she recommends a step-by-step approach: complete one part, then gradually connect and expand. For example, first integrate GitHub repository info, then organize domain-specific data, and finally generate PowerPoint presentations.

OpenClaw currently has over 18,000 Skills.

However, she notes that more Skills are not always better, as many require specific environments to run properly, involving additional installation, configuration, or costs, increasing usage complexity.

Additionally, OpenClaw can be configured with “.MD” files to set usage habits and personality, allowing users to control its behavior and make it more like a personalized assistant.

Risk boundaries: Permission management and cost control

Security risks are a major reason why OpenClaw has not yet become widespread. The official team is continuously fixing security issues through updates.

Xiao Cui admits his core principle is “try not to install on main machines” to avoid potential security risks. Allowing a possibly uncontrolled “intern” into your core workspace is unwise.

Notably, in early February, the Cybersecurity Threat and Vulnerability Information Sharing Platform of the Ministry of Industry and Information Technology detected that some instances of OpenClaw (commonly called “lobster”) posed high security risks when misconfigured or left default. These could lead to cyberattacks, data leaks, and other security issues. OpenClaw is an open-source AI agent that integrates multi-channel communication and large language models to create persistent, autonomous AI assistants deployable locally. Because of its “trust boundary” ambiguity and capabilities like autonomous operation, system calls, and resource access, lacking proper permissions, auditing, and security measures can lead to privilege escalation, information leaks, and system control issues.

Users and organizations should thoroughly check exposure to public networks, permission settings, and credential management when deploying OpenClaw. They should disable unnecessary public access, improve authentication, access control, data encryption, and security auditing, and stay updated with official security notices and recommendations to mitigate risks.

A notable incident involved a Meta executive nearly losing all important emails due to “lobster.” Peter Steinberger, known as the “father of lobster,” has emphasized privacy and data security, advocating for running agents locally.

Lynn states she grants full permissions only because her computer contains no sensitive data. Still, she emphasizes that while OpenClaw can reduce “hallucinations” via online searches, it may still provide outdated information.

Given the high permissions required, Steinberger advises developers to treat OpenClaw as a skill to practice, building intuition and defenses during use.

He said, “I can’t prevent users from misuse,” only helping them avoid shooting themselves in the foot.

On March 8, local time, OpenClaw’s latest iteration fixed 12 security issues.

Beyond security, resource consumption is an invisible cost of “shrimp farming.” Lynn recommends purchasing official Coding Plans from various vendors to avoid constantly worrying about token costs. She also warns that users may inadvertently set scheduled tasks that run long-term, causing token expenses to spike, memory overload, or disk space filled with logs.

Rational shrimp farming: essential for everyone or a transitional phase?

Regarding whether everyone should “farm a shrimp,” industry opinions are becoming more rational. Lynn believes OpenClaw is better suited for users eager to explore new tech trends and frameworks.

“Many now feel that not participating in this wave makes them outdated,” she said. “But after talking with many friends, I found that for them, ‘lobster’ doesn’t significantly improve efficiency over existing tools. Many tasks ‘lobster’ can do, other tools can also handle.”

She sees the core value of OpenClaw as proposing a new framework: “In the future, such capabilities might be integrated into wearable devices or other smart terminals. End users may not need to operate OpenClaw directly but will use its abilities through more natural interactions.”

Previously, nearly a thousand developers and AI enthusiasts gathered at Tencent Building, completing cloud deployment of OpenClaw with the help of Tencent Cloud engineers.

“Shrimp farming” is not just a tech experiment but an exploration of human-machine collaboration boundaries. In 2026, as we embrace the efficiency revolution brought by OpenClaw, remember to buckle up, experiment on idle machines, authorize cautiously, and rely on human wisdom. Only when technology ceases to be a burden and becomes a true invisible assistant can “shrimp farming” be considered truly successful.

(Source: The Paper)