OpenClaw, a crayfish-themed platform, has enriched model companies and cloud vendors. Will it become the next big thing in the AI ​​era?

Author: Frank, PANews

OpenClaw, which has become a global AI phenomenon, has developed its own "Chinese characteristics".

On March 6, 2026, an unusual scene unfolded in front of the Tencent Building in Nanshan District, Shenzhen: Tencent Cloud engineers set up a booth at the headquarters entrance to provide free installation services for OpenClaw to passing developers and AI enthusiasts.

This open-source AI agent framework, nicknamed "Crayfish" because its icon resembles a crayfish, is entering the public eye in a new kind of "ground promotion" approach.

In fact, the nationwide craze for "crayfish" has made it one of the fastest-growing non-aggregated software projects in the history of the code hosting platform GitHub, with the number of stars exceeding 250,000 in just a few months, surpassing established open source projects such as Linux and React.

Meanwhile, Tencent Cloud, Alibaba Cloud, JD Cloud, Volcano Engine, and Baidu Smart Cloud have all launched one-click deployment services, and an industry called "OpenClaw installation" has quietly emerged, charging between 100 and 500 yuan for remote installation. Some people claim to have earned 260,000 yuan in just a few days from installation services alone.

When a free, open-source tool needs to be promoted through "ground sales," and when a product that claims "everyone can have an AI assistant" spawns a business of packaging products costing hundreds of yuan, is this frenzy the prelude to the AI ​​Agent era, or just another bandwagon effect destined to end?

If we turn back the clock 20 years, the story of a product might offer some insights; it was called Xiaolingtong (Little Smart).

OpenClaw is indeed a good tool, but it's not "Jarvis".

Before discussing the fate of OpenClaw, it is necessary to recognize one fact: it is indeed an advanced product.

As an open-source AI agent framework, OpenClaw has accomplished something previously only achievable by a select few: connecting the capabilities of large language models (such as Claude, GPT-4, and DeepSeek) to everyday tools like WeChat, Telegram, DingTalk, and Lark through a unified interface. It's not just a simple chatbot, but a "digital employee" capable of browsing web pages, executing system commands, managing files, and writing code. As of March 2026, OpenClaw had reached 1.5 million weekly downloads on npm, and its plugin marketplace, ClawHub, boasted over 5,700 community-built skill sets with over 1,000 active contributors.

These data are sufficient to prove that OpenClaw has indeed hit the market's pain point. Just as the emergence of the PHS phone in 1998 allowed ordinary working-class people to use a "wireless phone" for the first time, OpenClaw has also given a large number of non-developers their first "AI assistant that can do things." The value of this market education should not be ignored.

However, from the perspective of an ordinary user, OpenClaw is still quite far from the AI ​​butler "Jarvis" from Marvel that people imagine.

First, there's the barrier to entry in terms of installation and usage. Deploying OpenClaw requires a Node.js environment, command-line operations, and API key configuration, which is almost an insurmountable obstacle for users without a technical background. This is also the fundamental reason why the contract manufacturing industry can exist.

More concerning are the hidden costs: some users have reported that the installation and debugging process alone consumed over $250 in API fees without producing any useful results. Even with successful deployment, the monthly token fees for heavy users can reach $100 to $1500. Behind the words "free and open source" lies a hefty hidden computing power bill. For those without AI experience, this could easily turn a simple task into a money pit. Therefore, money-saving guides have even appeared on the market, teaching users how to conserve token usage.

Secondly, there are security and stability concerns. Since 2026, several high-risk vulnerabilities have been disclosed in OpenClaw: CVE-2026-25253 allows remote code execution via malicious links, CVE-2026-25157 involves operating system command injection, and the "ClawJacked" flaw allows malicious websites to hijack local AI agents via WebSocket.

Because OpenClaw requires extremely high system privileges (reading and writing files, executing shell commands, controlling browsers, and capturing screenshots), an attack could have catastrophic consequences. A widely circulated case involves a security director at Meta whose imprecise instructions while using a similar AI agent led to the AI ​​mistakenly deleting hundreds of work emails. China's Ministry of Industry and Information Technology has also issued a security warning, reminding users to be wary of the potential risks of OpenClaw.

Furthermore, OpenClaw's performance is far less smooth than shown in the demo videos when handling complex tasks. Multi-level nested tasks can cause large models to enter infinite loops, and frequent API calls can easily trigger rate limiting mechanisms, leading to task interruptions. One user who tried using OpenClaw to automate daily office processes summarized their experience as follows: "I installed OpenClaw, spent the whole night fiddling with it. All the APIs were exhausted, and I didn't get anything done."

This sentence evokes a strong sense of déjà vu, much like that popular joke about mobile phones from 20 years ago: "Holding a mobile phone, standing in the wind and rain; switching from left hand to right hand, but still can't get through."

In terms of product maturity, today's OpenClaw is more like an "AI that needs to be served" rather than an "AI that serves you".

As a developer with over two years of Vibe coding experience, the PANews author recently attempted to deploy a "little crayfish" (a virtual machine), but the experience was extremely poor. Installing Skill and Connecting Channls alone took half a day, and all it could do was check the weather and mark events. For deeper programming, Cursor or Antigravity offer more control, directness, and stability. As for the automation touted by social media, it can be easily achieved through API-based large-scale model + program approach, regardless of cost or controllability.

Who is driving this frenzy?

If OpenClaw's product capabilities can only be rated as "acceptable," then why has it achieved such phenomenal popularity?

The answer may not lie in the product itself, but in the economic implications behind this frenzy.

The most direct beneficiaries are large model companies. OpenClaw is essentially a "token burner," with each task execution involving intensive calls to the large language model API. The token consumption of an OpenClaw Agent far exceeds that of traditional conversational AI chatbots, which is undoubtedly a godsend for large model companies urgently needing a "user growth story." China's large models and cloud services have also gained popularity due to their high cost-effectiveness, directly enabling the export of tokens overseas.

API packages from some major model manufacturers were snapped up, not because of a shortage of supply, but because OpenClaw created an unprecedented demand density.

Following closely behind are cloud providers. While OpenClaw emphasizes "on-premises deployment" to protect privacy, for most ordinary users, purchasing a cloud server to run OpenClaw is a more realistic option. Tencent Cloud, Alibaba Cloud, JD Cloud, Volcano Engine, Baidu AI Cloud—almost all major cloud providers in China launched one-click deployment services for OpenClaw immediately. Alibaba Cloud even launched a "Coding Plan AI Coding Package" specifically for OpenClaw users, covering the API demand spurred by OpenClaw with a fixed monthly fee.

On Tencent Cloud's Light Application Servers, the number of OpenClaw users has exceeded 100,000. Tencent's decision to offer free installation in front of its headquarters, ostensibly a charitable act, is actually a targeted user acquisition strategy. It allows you to install OpenClaw for free, but you'll need to continue paying to use Tencent Cloud's servers to run it.

This logic is exactly the same as the low-cost strategy of China Telecom during the era of PHS (Personal Handyphone System): attract users with low barriers to entry and then retain them with continuous service fees.

Another easily overlooked underlying factor is hardware demand. OpenClaw's recommendation for local deployment directly boosted demand for computing power equipment. The overseas installation platform SetupClaw charges between $3,000 and $6,000, often including "recommendations" for specific hardware configurations. The operating logic of this industry chain is structurally highly similar to the story of how the PHS (Personal Handyphone System) spurred base station construction and drove the entire telecommunications equipment industry chain 20 years ago.

Looking back at the history of PHS (Personal Handyphone System), its rapid rise to popularity in the Chinese market was not primarily due to its superior product capabilities, but rather because China Telecom, lacking a mobile communication license at the time, urgently needed to expand its revenue streams through this "quasi-mobile" service. The driving force came from the company's commercial interests, rather than genuine consumer demand.

The same applies to OpenClaw today: large model companies need increased usage, cloud vendors need server sales, and hardware manufacturers need to ship computing equipment. When a product's popularity comes more from the supply side than from the demand side, its prosperity is often fragile.

The ultimate form of AI automation: integration, not assembly.

If OpenClaw is just a transitional product, then what should a true AI Agent look like?

The answer is emerging. 2026 is widely regarded by the industry as the "Year Zero of AI-Native Phones," with many tech giants integrating AI Agent capabilities directly into operating systems and hardware devices, rather than requiring users to install a third-party framework.

ByteDance, in collaboration with mobile phone manufacturers such as Vivo, launched "Doubao Mobile Assistant," which deeply embeds AI Agent capabilities into the underlying layer of the mobile operating system. Users can simply press the side button to allow AI to complete complex tasks across applications, such as "comparing prices and placing orders across multiple platforms," ​​"automatically ordering food and hailing rides," and "integrating travel guides to generate itineraries." The entire process is executed automatically in the background, requiring no framework installation or API configuration.

On March 7th, Xiaomi announced that Xiaomi miclaw, built on its self-developed MiMo big data model, has begun closed testing. The goal is to deeply integrate it into the phone's underlying system, calling upon over 50 system tools, and ultimately controlling more than 1 billion Mi Home smart devices. Overseas, Windows Copilot, Apple Intelligence, and Gemini in Android are following the same path.

IDC predicts that shipments of next-generation AI smartphones in the Chinese market will reach 147 million units in 2026, accounting for more than half for the first time, reaching 53%.

This means that AI Agent is transforming from a geek toy that users need to "assemble" into a "ready-to-use" system-level capability.

Comparing OpenClaw with these native AI products side-by-side reveals the differences: OpenClaw requires users to set up the framework, configure the large model API, and connect to each platform one by one, essentially making it a "universal adapter"; while the Agent in native AI phones and operating systems is a built-in capability that is ready to use out of the box, requires no installation or configuration, and has its security fully guaranteed by the system manufacturer.

This comparison almost perfectly mirrors the relationship between PHS (Personal Handyphone System) and 3G phones. PHS was phased out not because people no longer needed to make calls, but because 3G phones offered better, more portable, and wider-coverage calling capabilities. Similarly, OpenClaw's potential marginalization in the future won't be due to a lack of need for AI agents, but because natively integrated AI agents will deliver an experience that OpenClaw can never match.

Echoes of History: From a Young Person's Perspective on OpenClaw's Fate

Here, it is necessary to briefly review the life trajectory of PHS (Personal Handyphone System) to more clearly understand why OpenClaw is called the PHS of the AI ​​era.

The PHS (Personal Handyphone System) technology originated in Japan and was introduced to China by UTStarcom in 1998. Its essence was not mobile communication technology, but rather a wireless extension of fixed-line telephones, using microcell base stations to wirelessly connect user terminals to the local fixed-line network. Its rapid rise to popularity had only one core reason: affordability. In an era when mobile phone call charges were high and both incoming and outgoing calls were billed, PHS's one-way billing (free incoming calls) and low monthly fee allowed many working-class people to use "wireless phones" for the first time, hence its nickname "the poor man's phone."

In October 2006, the number of PHS (Personal Handyphone System) users in mainland China reached a historical peak of 93.41 million.

However, technical flaws persisted. Poor signal coverage, lack of nationwide roaming support, and potential disconnections at speeds exceeding 40 kilometers per hour: "Holding a PHS phone, standing in the wind and rain" wasn't a joke, but a real user experience. More fatally, as mobile phone tariffs continued to decrease and 3G technology matured, PHS's only price advantage was gradually eliminated. In 2009, the Ministry of Industry and Information Technology required PHS to complete frequency clearing and network decommissioning by the end of 2011. In 2014, PHS base stations in mainland China were gradually shut down, marking the end of its 16-year history.

Applying the story of PHS (Personal Handyphone System) to OpenClaw, three implications are worth considering.

First, the reason why PHS (Personal Handyphone System) became popular was not because it was good, but because there were no better options at the time. During a period when 3G phones were not yet widespread and mobile phone tariffs were high, PHS provided a "good enough and cheap" alternative. OpenClaw faces a strikingly similar market environment today: native AI agents are not yet mature, official agent products from major model vendors are still being refined, and operating system-level AI integration is just beginning. In this vacuum, OpenClaw filled the gap with its "free, open-source, and customizable" approach. However, "filling a gap" and "defining the future" are two different things.

Secondly, the decline of PHS (Personal Handyphone System) wasn't due to its deterioration, but rather the arrival of better technology. PHS attempted self-evolution: launching an MMS version and trying to expand coverage. However, these improvements ultimately couldn't bridge the fundamental gap between its underlying architecture and true mobile communication. Similarly, OpenClaw can continue to iterate, add more features, and optimize deployment processes, but its essence as a "middleware framework" remains unchanged. When Doubao Mobile Assistant allows users to perform cross-application operations with a single click, when Xiaomi miclaw can directly control all smart devices in the home, and when Apple Intelligence becomes a standard feature of the iPhone, a third-party agent framework that requires user installation, configuration, and maintenance, like PHS in the 3G era, hasn't deteriorated; the world has changed.

Third, when China Telecom launched PHS (Personal Handyphone System), it wasn't because it represented the future, but because it could generate revenue in the present. China Telecom lacked a mobile communication license, and PHS served as a roundabout way to enter the market. Today, cloud vendors' investment in OpenClaw follows the same logic: not because OpenClaw represents the future of AI, but because it can sell cloud servers, drive token consumption, and acquire users in the present. When better AI agent products emerge, these vendors will switch battlefields with the same decisive speed as China Telecom's shift to 3G.

However, any analogy has its limitations. It took 16 years for PHS (Personal Handyphone System) to disappear, while the story of OpenClaw is just beginning. The iteration speed of AI technology far outpaces the generational shifts in communication technologies. This means that the window of opportunity for OpenClaw to go from "being popular" to "replacing" PHS may be much shorter than that of PHS. But it also means that the value it created for the industry during this window should not be entirely negated. It allowed hundreds of thousands of non-technical users to experience the possibilities of AI agents for the first time; its open-source ecosystem provided the community with a low-cost experimental platform; and the security, cost, and stability issues it exposed provided valuable lessons for those who followed.

But history doesn't change course because of popularity. At its peak, PHS (Personal Handyphone System) had 93.41 million users, but scale couldn't withstand the tide of technological advancement. OpenClaw boasts 250,000 GitHub stars, but the number of stars is never a measure of a product's longevity. When AI capabilities are truly integrated into the phones, computers, and operating systems we use every day, when "AI assistants" are no longer software that needs to be installed but rather a ubiquitous infrastructure like Wi-Fi, few people will miss the "Little Lobster" (a popular Chinese mobile phone brand) that took an entire night to install.

In this nationwide craze for installing crayfish-themed apps, what's truly worth considering isn't what OpenClaw can do today, but whether we're ready to embrace a truly AI-native era when it's no longer needed.

After all, the PHS phone taught us a simple truth: in the long race of technology, the one that runs to the end is always the product that you don't have to work hard to adapt to.