In March, people were still queuing up for proxy installations; by April, many had already quit. The hype around OpenClaw resembles a rollercoaster ride.

On March 3, OpenClaw's GitHub stars surpassed React, achieving in less than 60 days what the latter took a decade to accumulate. Its WeChat Index soared to 165 million, with narratives of "24-hour digital employees" sweeping through China's AI circles. "National Shrimp Farming" became another cultural phenomenon following DeepSeek. By March 15, cumulative monthly token usage reached 10.4T, making it the world's most-used AI application.

But a month later, the tide suddenly turned.

In April, OpenClaw's traffic plummeted to 14.2 million, a 50.67% month-on-month drop. Tencent's QClaw saw a 99.19% decline, while domestic "lobster" products like Kimi Claw and CoPaw also experienced sharp drops. Developer forums began to see posts declaring: "OpenClaw is dead."

Amid this retreat, two notable events occurred. In April, Anthropic announced that subscriptions would no longer support OpenClaw, giving users just one day's notice. In early May, OpenAI's Sam Altman tweeted in the early morning, announcing that ChatGPT users could now log into OpenClaw—"happy lobstering."

One company kicked the lobster out, while another opened the back door to welcome it in overnight. To some extent, OpenClaw's decline may not mark the end of the Agent craze but rather the beginning of the AI industry's reflection on how AI Agents should operate, who should control them, and how they should be monetized.

I. A Foreseeable Retreat

To understand its decline, we must first consider why it became so popular.

Most analyses attribute OpenClaw's success to its product strength, such as supporting over 50 integrations, compatibility with Claude, GPT-4o, Gemini, and DeepSeek, and attracting enterprise users—Tencent even built a platform directly on top of it. While these factors are valid, they only tell half the story.

The real driver was access rather than capability. OpenClaw made it cost-effective to use powerful models at scale.

Users discovered that OpenClaw could leverage Claude Pro or Max users' login credentials, simulating Claude Code client requests to access Claude. This meant users no longer needed to pay per token for each API call but could instead rely on fixed monthly subscriptions for high-frequency Agent usage at a lower cost.

At its core, this was a silent form of arbitrage.

The subscription price for Claude Max is $200 per month. Industry analysts estimate that running the same workload via API costs more than five times as much. In other words, Anthropic was subsidizing hundreds of dollars per month for each heavy OpenClaw user.

This near-free computational power fueled OpenClaw's viral spread, launching the growth myth of "National Shrimp Farming."

However, in April of this year, Anthropic sent an email notifying users that Claude subscriptions would no longer cover high-intensity third-party Agent usage like OpenClaw, with the new rule taking effect immediately, leaving users just one day to adapt.

The impact of this email on the OpenClaw community was akin to someone pulling the plug at a raging party.

Simultaneously, users deploying OpenClaw at scale began to encounter its practical limitations. For instance, Agents were fragile, failing unpredictably in multi-step tasks, with inconsistent reliability across different workflows and environments.

More importantly, for most non-technical users, setup and maintenance were not straightforward.

During OpenClaw's peak popularity, domestic markets even saw the emergence of proxy installation services—charging fees ranging from tens to hundreds of yuan to help users with deployment, API configuration, and permission setup. These services were in high demand, indicating that a tool requiring paid installation and remote debugging was far from being a truly mass-market product.

Additionally, another crisis quietly eroded user confidence.

Unlike ordinary AI chat tools, OpenClaw didn't just "talk"—it "acted." It could open folders, send emails, execute commands, and invoke various software, granting it far greater access to users' computers than they realized.

This looked cool in demo videos. But in reality, the stakes were high. If an ordinary AI chat tool hallucinated, it might give you a wrong answer; if OpenClaw hallucinated, it could delete your files, leak your account credentials, or execute destructive commands without your knowledge.

Worse, many users had already connected OpenClaw to their most sensitive data—work emails, cloud storage, even corporate server permissions—without fully understanding the risks.

Thus, discussions around OpenClaw gradually shifted from "efficiency revolution" to "risk out of control."

II. Where Did the Users Go?

After OpenClaw's hype faded, where did its users go?

Broadly speaking, they fell into three categories: heavy developers migrating to Hermes, professional users turning to Claude Code, and enterprise users adopting vertical industry Agents.

Hermes has emerged as the primary destination for displaced OpenClaw users, a fact supported by data.

On May 10, Hermes surpassed OpenClaw with 224 billion tokens consumed in a single day, topping OpenRouter's global daily inference rankings for the first time since OpenClaw's release.

A new saying began circulating in the community: "Lobster farming is out; horse breeding is in."

Hermes includes a built-in one-click migration tool designed specifically for OpenClaw users, automatically importing configurations, memories, skill packages, and API keys.

But the migration tool was just the entry point; what truly retained users was the difference in design philosophy.

OpenClaw started each task from the same baseline, ending upon completion and restarting anew next time. Hermes incorporates a learning loop, automatically generating reusable Skill files after task completion, building persistent understanding across sessions. Unlike most stateless AI tools, Hermes operates on the opposite logic. After performing a competitive analysis once, users can simply say "do it like last time" next time, without needing to redescribe the process and requirements.

The deeper the usage, the higher the migration cost—this is Hermes' underlying logic for retaining heavy users. Currently, Hermes has accumulated 110,000 GitHub stars, with community contributions focused on skill quality and learning improvements rather than functional expansion. This contrasts sharply with OpenClaw's early ecosystem expansion strategy of "doing everything," resulting in a more professional and higher-frequency user base.

Besides Hermes, another group of professional developers moved to Claude Code.

This is a command-line programming Agent directly provided by Anthropic, which doesn't attempt to manage your emails, schedule, or desktop but focuses solely on completing complex tasks at the code level, with controlled execution environments and predictable behavior.

This diversion (user diversion ) reveals a genuine value split within the developer community: OpenClaw prioritizes openness and unrestricted capabilities, while commercial systems prioritize reliability and controlled execution. As AI programming tools mature, many developers realize they don't need highly autonomous but unpredictable Agents—they need stable systems.

In other words, "let AI do everything for me" and "let AI perform better in specific scenarios" represent two entirely different user needs with varying intensities. OpenClaw's decline has reseparated these groups.

Another group of users chose specialized industry tools instead of Hermes or Claude Code.

Products like DingTalk Wukong and WindClaw follow the opposite logic of OpenClaw—they don't aim to do everything but embed Agent capabilities into existing enterprise workflows to handle highly repetitive tasks like approvals, reporting, and investment research briefings.

However, one fact is easily overlooked: the largest group of users during the retreat didn't migrate to any product.

Followers returned to Doubao and ChatGPT, while casual users never found genuine use cases. For them, the Agent category simply wasn't suitable at this stage because the tools were too complex, costs too high, and needs too infrequent.

III. Agents Begin to Wear "Shackles"

As OpenClaw's hype faded, another easily overlooked event occurred.

On April 4, Anthropic sent users a ban email; shortly after, in May, OpenAI CEO Sam Altman tweeted on X: "ChatGPT users can now log into OpenClaw. Happy lobstering."

One company cut off connections, while another actively opened an entry point. This to some extent (to some extent) exposes the emerging ideological divide within the AI industry.

For the past year, the AI industry's most captivating vision for Agents has been the "digital employee."

Products like OpenClaw, AutoGPT, and Devin depicted a future where AI completes tasks independently like humans: reading emails, manipulating tools, writing code, collaborating across software, and autonomously executing complex workflows.

But once these Agents entered real-world scenarios, problems began to surface. The more autonomous the Agent, the more prone it was to out of control (losing control); the longer the task chain, the more likely it was to deviate from goals; the higher the permissions, the more hesitant enterprises were to grant true autonomy.

Many users later realized that what they truly needed wasn't an AI to "work for them" but an AI to "assist them in their work."

Thus, the industry began a noticeable shift.

The OpenClaw generation believed: the more open, the more human-like, and the more autonomous, the stronger. But the new generation of Agent products began to reconsider: restricting permissions, narrowing scenarios, and retaining human confirmation might actually make implementation easier.

Hermes now emphasizes long-term memory and skill reuse, abandoning the pursuit of "doing everything" and instead enabling Agents to continuously learn user habits within fixed workflows. Claude Code further narrows its capability boundaries. It doesn't attempt to control the entire computer environment for users but limits its abilities to coding scenarios, making execution processes more controllable and stable.

Thus, the entire industry began shifting from "Autonomous Agents" back to "Human-in-the-loop" systems—meaning AI no longer tries to fully replace humans but returns to assisting them.

This change is even more pronounced among major tech companies.

Microsoft began repackaging OpenClaw-style Agent capabilities into its Copilot ecosystem; NVIDIA launched NemoClaw, focusing on enterprise-grade security and permission controls; domestic vendors transformed "lobsters" into vertical products like DingTalk Wukong and WindClaw, embedding them into specific approval, reporting, and investment research processes.

They all share one commonality: no longer pursuing "doing everything" but embedding Agents into specific workflows. Because what enterprises truly need isn't an AI adventurer that randomly executes commands and might accidentally delete databases but a stable, deterministic, and auditable digital system.

To some extent, OpenClaw resembles a typical "custom-built PC."

Users source their own models, configure environments, tune APIs, and fix bugs themselves. It's powerful yet extremely fragile. Geeks revel in this process, but ordinary users struggle to endure it long-term.

This resembles the personal computing era before the iPhone's emergence. During that phase, tech enthusiasts loved "building their own PCs," while average users just wanted to press the power button. The Agent industry now stands at a similar crossroads.

Everyone recognizes that AI Agents will be the next major human-computer interaction paradigm. But most people don't want to "farm shrimp"—they want a pre-packaged product.

In just a few months, OpenClaw has shown the entire industry that the future of AI Agents may not belong to the freest systems but rather to those that are most stable, secure, and user-friendly for ordinary people.