As the Lobster Tide Recedes, Harnesses Rise
By Mengyao, reporting from Aofeisi Temple
It hasn’t even been that long—how does it feel like OpenClaw never existed? … (confused.jpg)
Recently, the collective mourning for OpenClaw among users on X has left ordinary netizens sighing with emotion—and Peter, the father of the Lobster, in tears…
The thing is—
Come on, this thing was still insanely popular just a few months ago!!!
Just a few months ago, the entire AI community was celebrating its own “JARVIS moment.”
Everyone was raising lobsters, scrambling to buy Mac minis, and paying for on-site installations. Overnight, everyone on social media seemed to become a cybernetic lobster farmer. The spectacle was so huge that it became a genuine viral phenomenon—and it still feels like yesterday.
Yet barely half a year later, everyone has collectively turned their attention to Claude Code, Codex, Skills, and Harnesses…
Fine, fine. The AI world now measures the shelf life of even its “tears of the times” in months.
OpenClaw: So love can disappear, right?
From “JARVIS moment” to tears of the times—OpenClaw needed only a few months
The world can be downright surreal sometimes.
Many things that later become known as “tears of the times” were often so wildly popular when they first appeared that they seemed destined to rule the next era.
OpenClaw was a classic example. (doge)
The main reason it went viral on such a massive scale was that it genuinely upended everyone’s understanding of AI…
At the time, people were still figuring out how to use various AI products effectively. Then Peter, the father of the Lobster, suddenly unveiled the work of the century: OpenClaw—
An AI agent that lived on your computer 24/7, waiting for your commands at any moment.
You could install it directly on a Mac, Windows PC, or even a Raspberry Pi, then connect it to models such as Claude, GPT, Gemini, and DeepSeek.
After that, you could hand over access to your email, browser, terminal, calendar, messaging apps, and more.
From there, you didn’t even need to stay in front of your computer.
Send it a message in Discord or Telegram, and it could read emails, look up information, write code, run scripts, and even install new tools for itself.
At the time, this experience hit everyone’s most basic fantasy about AI agents squarely on the mark:
Whoa, someone actually built an open-source JARVIS!!!
And then the entire AI community started celebrating.
OpenClaw quickly went from a developer project to a nationwide “lobster-raising movement.”
For a while, buying a Mac mini, installing Clawdbot, and naming your own agent became the AI community’s new form of cyber-decoration.
Huaqiangbei in Shenzhen even experienced a temporary shortage of Mac minis, with the prices of some models rising by as much as RMB 600 for in-stock units.
The major tech companies soon joined in as well.
Companies such as Tencent and Baidu held various offline OpenClaw installation events.
Dozens of programmers set up booths on-site, providing hands-on help to hundreds of people installing OpenClaw. Online, services offering home installation of a lobster for RMB 499 were also competing fiercely.
After that, one-click deployment solutions began appearing everywhere. “Lobster-raising guides,” zero-barrier versions, and all kinds of XX Claw products flooded the market.
On GitHub, OpenClaw’s Star count shifted from a steady climb into full-on “pulling a scallion straight out of dry ground” mode.
From its launch at the end of November 2025 to March 3, 2026—roughly 100 days—OpenClaw shot past 250,000 Stars, overtaking React in the process.
React took more than a decade to reach that scale. OpenClaw did it in just over three months.
By the end of March, the Lobster’s Star count had surged past 330,000, briefly making it one of the fastest-growing open-source projects in GitHub history. Some statistics even showed that it had gained more than 34,000 Stars in a single 48-hour period at its peak.
The world is just that ruthless. It shows JARVIS no mercy.
Now, only a few months have passed since OpenClaw appeared out of nowhere.
When I searched for OpenClaw on social media again, it genuinely felt a little like archaeology…
The Lobster that once stunned, dazzled, and left the entire internet slumped in its chair had now become the subject of collective mourning whenever it was mentioned again—
Netizen A: OpenClaw was supposed to rule the world, but now nobody even talks about it anymore. (lol)
Netizen B: What happened to OpenClaw? I almost never see anyone mention the Lobster anymore…
Netizen C went even further: The Lobster is totally dead, right? Nobody seems to be discussing the product anymore. I guess it’s done for.
All you can say is that OpenClaw peaked right out of the gate. Once the peak passed, the Lobster effectively disappeared…
The Lobster isn’t dead, but everyone has started using “official substitutes”
At this point, we should give OpenClaw its due: the Lobster isn’t actually dead!!
According to the GitHub data, it’s doing quite well~
By August, OpenClaw’s Star count had climbed past 380,000, with more than 80,000 Forks. Related products on X were still being updated, and people were still developing, tinkering with, and raising lobsters.
More accurately, OpenClaw hasn’t really stopped being popular—it has simply “softened the rain and nourished the earth”—
Because the “ultra-approachable” versions of the Lobster have swum into the ponds of all the major vendors and begun multiplying like crazy…
Around March this year, Tencent launched WorkBuddy and QClaw in succession. Both emphasized one-click deployment and could be controlled remotely through WeChat.
Baidu released its DuMate+ Lobster family bundle, offering a desktop lobster, cloud lobster, mobile lobster, secure lobster, and home lobster—all in one pot.
NetEase Youdao also launched LobsterAI, supporting local personal agents, cross-platform operation, Skills, MCP, and remote control through IM.
Zhipu launched AutoClaw, which supports one-click local deployment of OpenClaw, comes with preinstalled Skills, and integrates with Feishu. It later began competing in the “self-evolution” space as well.
Volcano Engine launched ArkClaw, which likewise supports 24/7 operation and later expanded toward Runtime and multi-agent collaboration.
……
According to incomplete statistics, domestic vendors alone have launched more than 30 Lobster derivatives.
In other words, although the name OpenClaw is gradually fading, low-barrier, affordable lobster fry has already been released into everyone’s ponds…
And that is only the first layer of change.
Because what truly brought the era of “everyone raising lobsters” to an end was the AI community moving on to the next layer of competition—
Harnesses now rule.
This says a great deal about what OpenClaw has gone through over the past few months.
Remember? At the height of the Lobster craze, its most captivating feature was that it could “assemble” almost anything on its own.
It chose the model, configured its own permissions, connected to email, browsers, and terminals, and let you keep stuffing Skills into it layer by layer.
But a few months later, “official” offerings such as Claude Code and Codex had packaged this entire experience and delivered it as a finished product—
When OpenAI updated its Agents SDK this year, it explicitly placed capabilities such as file operations, command execution, long-running tasks, and sandbox execution into a model-native harness. Codex itself has also developed a complete Harness of its own.
How to provide models with context, memory, tools, sandboxes, permissions, and workflows so they can work steadily and continuously has become the central topic of discussion.
All you can say is: every generation has its own god of the era.
Of course, the shift in the spotlight cannot be blamed entirely on the AI community’s short attention span.
I don’t know whether everyone remembers this, but the features that made OpenClaw so impressive in the first place also came with several very real “side effects.”
The first was simple: it burned money.
At the time, “raising lobsters” fueled a wave of Token anxiety. An agent that stayed online 24/7 would constantly maintain context, call models, call tools, and retry failed operations. It really did consume a lot of Tokens.
So, for many people, the lobster-raising experience went something like this—
Day 1: This is incredible—isn’t this an open-source JARVIS!?
Day 3: Huh? It can do that too?
Day 7: Someone help me—where did all my Tokens go?
There was also another keyword that had followed OpenClaw since birth: security.
One of the most classic examples was when Meta’s head of security and coordination granted OpenClaw access to an email account.
And then, well, the Lobster went on a rampage and started deleting emails—and for a while, it couldn’t be stopped at all…
What made the incident so frightening was that it exposed the core contradiction of personal agents: the more you want them to act like JARVIS, the more permissions you have to give them; and the greater their permissions, the higher the cost of losing control even once.
So, looking back at OpenClaw, we can see that it didn’t really lose all that badly.
The Lobster seems to have demonstrated in advance all the problems that the next generation of agents would truly need to solve—
How to give AI hands and feet without letting it reach where it shouldn’t; how to keep it working 24/7 without burning money 24/7; how to give it enough freedom while still being able to pull it back when something goes wrong.
And the things that Claude Code, Codex, and an increasing number of Harnesses are frantically filling in today are precisely these gaps.
So, in a sense, the greatest legacy left by the Lobster may have quietly shifted from “everyone should raise an OpenClaw” to:
Everyone should start figuring out how to build a more reliable OpenClaw!!
What is the father of the Lobster doing now? Building a factory for the next generation of JARVIS
Which brings us to the next question.
With the Lobster nearly reduced to a tear of the times, where is Peter, the father of the Lobster, making his fortune now?
The answer is a little cinematic: he went to OpenAI to build the next generation of agents.
In February this year, Peter Steinberger officially joined OpenAI. His focus was clear: continuing to work on the thing he was most obsessed with—personal AI agents.
After joining OpenAI, Peter quickly immersed himself in work related to Codex and agents.
While building a team, he continued exploring how to make agents work autonomously for longer periods and with greater stability.
Recently, he has repeatedly mentioned the concept of a Software Factory in interviews and talks—
A group of agents writes code, tests it, fixes bugs, and continues iterating, while people are mainly responsible for setting goals and signing off on the results. He has even tried letting Codex run continuously on a task for an entire week.
And just in the past couple of days, Peter appeared in another rather surreal piece of news—
Together with Dropbox co-founder Drew Houston, he each contributed $1 million to become a new Founding Patron of the Omacom Foundation.
What does this foundation do?
Seriously investigate how to stop the “Year of the Linux Desktop” from being postponed year after year. (not really)
More seriously, it uses funding to support infrastructure, open-source projects, and developers, with the goal of truly bringing the Linux desktop to a broader audience.
Yes, raising one lobster is no longer enough for the father of the Lobster.
Here, OpenClaw is almost a cybernetic artifact.
Over there, Peter has already gone off to build a factory for the “next generation of JARVIS.”
OpenClaw: No matter what, at least we loved each other. (doge