The "Father of Lobsters" is trapped in lobsters.

"Lobster" finally has new updates, but the "lobster craze" has long faded away.

On August 31, OpenClaw released the v2026.8.1 version, which the official calls "OpenClaw 2.0".

The usage threshold is no longer as high as it used to be, and OpenClaw 2.0 has added a large number of new features, such as memory organization, multi-person collaboration, and many more.

However, the traffic of OpenClaw's GitHub page was cut by half from its peak of about 28.8 million in April to 14.2 million, and before this update, it had dropped below 10 million.

In February and March, OpenClaw averaged 5,900 new GitHub stars per day. In the past two weeks, however, the number of new stars added each week is about 800. In contrast, Hermes Agent gains nearly 9,000 new stars per week, a difference of more than 10 times.

In terms of weekly NPM downloads, the figure peaked in mid-March at over 1.5 million, and has now fallen back to the level before OpenClaw went viral.

Worse still, Crabbox, the new product Steinberg is working full-time to maintain, has barely attracted any attention. Its total number of stars on GitHub is even less than the number of new stars OpenClaw gains per day after the craze died down.

The "Father of Lobsters" seems to be trapped in his own lobster creation.

How the Lobster Was Raised

OpenClaw 2.0 is the largest update since the birth of "the lobster", with 933 contributors participating, 569 of whom submitted code for OpenClaw for the first time. More than 16,000 Pull Requests have been merged in total, nearly half of the total merged volume in the project's history. Before the release, OpenClaw did not release a stable version for nearly 7 weeks.

In the 230 days before the "update pause", OpenClaw released 106 stable versions in one go, averaging an update in less than two days. Sometimes, right after you finished updating and were about to use it, the next update was already pushed out.

Steinberg said that the biggest problem with OpenClaw is that its underlying architecture is too outdated. So this time they spent two full months on refactoring, preferring to postpone the original scheduled launch time to make sure that old users would not break their configurations when upgrading.

The biggest problem solved by the 2.0 version is installation.

As mentioned at the beginning, in the 1.0 era, the threshold for "raising a lobster" was ridiculously high. You needed an API Key plus a bunch of command line configurations, and anyone with insufficient computer knowledge could basically say goodbye to using it.

Version 2.0 simplifies the initial installation to out-of-the-box use, supports direct access to existing ChatGPT and Claude subscriptions, also supports API Key and local models, and provides multiple login methods. The browser application has been completely rewritten, the original Overview page has been removed, and now it only takes 575 milliseconds from startup to entering the conversation.

More important than the interface is memory.

In the 1.0 era, OpenClaw would crash as soon as its context expanded, forcing users to re-explain the background every few days. Version 2.0 introduces the Active Memory mechanism, allowing the Agent to call relevant contexts from past conversations in eligible sessions.

It also adds the new Background Memory Consolidation feature, where the model automatically organizes long-term memory in the background, promotes information that is truly worth retaining to long-term storage, and retains source information and "Dream Diary".

The so-called Dream Diary refers to the process where the Agent organizes the memories left from its work during non-working hours. On this basis, the self-learning mechanism allows the Agent to precipitate reusable methods from completed tasks, automatically convert them into new Skills, and achieve a certain degree of self-evolution.

Version 2.0 also put the most effort into security.

When the Agent needs a key, it requests it through a masked prompt. The key value never enters the chat record, nor does it enter the model's context window, just like when you enter a password at an ATM, the machine only displays ****.

In addition, only one trust boundary is set for each Gateway, and any installation that exposes the instance to the public network without authentication will be blocked before the change takes effect.

The official also released a set of crowdsourced red team data: 272,000 attacks across 41 Agent scenarios, with the success standard being "performing harmful actions while deceiving the user". The lower the success rate, the safer the product. The success rate is only 0.5% when using Claude Opus 4.5, 1.0% for Sonnet 4.5, 1.3% for Haiku 4.5, and even the worst-performing Gemini 2.5 Pro only has an 8.5% success rate.

Version 2.0 also makes up for the shortboard in collaboration.

The newly introduced Shared Cloud Sessions allow multiple people to join the same task, transfer work with full context, and the permissions are divided into four levels: read, suggest, draft, and direct participation, similar to common collaboration software.

The OpenClaw official said that their own team uses this feature to develop OpenClaw.

The OpenClaw team said that initially they only wanted to simplify installation and redo the browser side, but as they worked on it, they found that to let non-technical users "raise lobsters", they had to completely rewrite the sessions, memory, permissions, keys, cloud execution, multi-Agent collaboration and team sharing all at once.

In other words, rather than deciding to release a 2.0 version, it is the workload of this update that naturally grew into what is now version 2.0.

But all the features launched in OpenClaw 2.0 have long been available in other Agent products on the market.

Almost all Agents on the market, except for DeepSeek's DSH, have no installation threshold at all. You just download the installation package to your computer, click on it, and you can install and use it.

At the same time, products like Codex and Claude Cowork have previously launched features similar to memory consolidation and active memory, and they also perform better at the security level.

Rather than saying that OpenClaw has finally updated to 2.0, it is more reasonable to ask why it took OpenClaw so long to get to 2.0?

Why Did the Lobster Start to Go Downhill?

Months after the lobster craze died down, Steinberg reviewed the experience on YC's startup podcast in August, saying: "Be careful what you wish for. I wasn't prepared for all this attention, and it almost crushed me."

The so-called "wish" refers to a post Steinberg once published on X, hoping that his project would be seen by the whole world and go viral.

Steinberg went on to say that his private phone number was leaked, and journalists would call him late at night using emergency penetration modes. He also complained that every few weeks, there would be an obituary-style report titled "OpenClaw is Dead", and he had lost count of how many of those he had read.

"One of them was the so-called OpenClaw killer, which turned out to be nothing but an uninstaller."

After joining OpenAI, OpenClaw was "no longer fun" for him.

Steinberg said he woke up in the morning and found he had two jobs: one was a job, the other was a mission. The worst symptom, he said, was that he stopped using OpenClaw himself.

"I could no longer happily add new features to the lobster, and instead became the person who fixed bugs, handled security incidents, and replied to user messages every day," Steinberg said.

At the end of August, in GitHub's official maintainer interview, Steinberg said: "I don't even call them Pull Requests. I call them 'prompt requests'."

A large number of contributors ran "automated software factories" and poured in hundreds of Pull Requests at once, all generated in batches by AI.

Steinberg said these Pull Requests completely distracted him, preventing him from focusing on product development, and he had to spend a lot of energy every day managing these Pull Requests.

That's not all. Another reason for OpenClaw's decline is the supply cutoff from upstream.

In February, Anthropic updated its terms, prohibiting the OAuth quota of Claude subscriptions from being connected to third-party tools like OpenClaw. On April 4, Anthropic officially sent an email notifying that Claude subscriptions would no longer cover usage through third-party tools such as OpenClaw. To continue using it, users would have to buy a 30% off usage pack or use the API.

The vast majority of users previously used Claude to "raise lobsters", and after Anthropic's move, the cost of using OpenClaw multiplied sharply.

Steinberg admitted on the podcast that he made a strategic mistake: "To make OpenClaw perform best, I deeply optimized the system for Anthropic's models, but this company ended up becoming our own competitor."

"They beat us at the most critical point, while we stayed where we were," he said. "Both of these mistakes are mine, I caused them." He also said a line that almost deconstructed the "lobster myth" by himself: "Eight months is not a long time in human time, but in AI time, it is equivalent to four years."

The concept of OpenClaw is very grand, but the technology is difficult to keep up. Back then, the Gateway process of OpenClaw could crash at any time with no automatic recovery; memory was only at the session level, and it would crash as soon as the context expanded; most of the Skills were written manually with uneven quality, and the maintenance team was mainly made up of part-time volunteers.

Then came the data that everyone could see: according to statistics from Feifan Research, OpenClaw's traffic was cut in half from its peak of about 28.8 million in April to 14.2 million.

In February, due to the high installation threshold of OpenClaw, the "on-site lobster installation" service on Xianyu became a new way for programmers to make money. By April, the "on-site lobster uninstallation" business began to rise. From installation to uninstallation, a complete closed loop was completed in just two months.

However, while the whole world was complaining about the lobster, the shadow of the lobster was everywhere in the market.

Competitors paid tribute to it while making targeted improvements. Nous Research open-sourced Hermes Agent at the end of February, tagged with self-evolution, automatic abstraction of tasks into reusable Skills, four-layer memory, built-in sandbox and dangerous command approval, and one-click curl|bash installation, which specifically solved all kinds of pain points of OpenClaw. It gained tens of thousands of stars in two months, and the community put forward the slogan "From raising lobsters to raising horses (the horse refers to Hermes Agent)".

Moon & Dark released KimiClaw, which is fully cloud-hosted with zero-threshold registration; Alibaba Cloud Tongyi launched CoPaw, with dual local and cloud deployment, and deep integration with DingTalk; Tencent has QClaw, ByteDance has Moltbook...

The Ignored "Cloud Lobster" — Crabbox

About two months after joining OpenAI, he developed a project called Crabbox, a general-purpose remote software testing and execution control platform.

Specifically, it will rent a disposable virtual machine from cloud vendors such as AWS, Hetzner, Azure, and GCP, use rsync to synchronize your local uncommitted changes (dirty checkout) to it, execute commands remotely, stream stdout/stderr back in real time, record operation evidence (logs, screenshots, videos, artifacts), and delete the machine to release resources after the task is finished.

What Crabbox does is actually very practical. After you finish writing code on your local computer, you need to run tests, builds, and AI inference. These tasks are not only performance-intensive, but also time-consuming and delay your work. So Crabbox will do these things for you on the rented cloud server.

To put it simply, it is a cloud lobster, which moves the whole process of "raising lobsters" to the sandbox in the cloud.

Compared with the original lobster, Crabbox is more suitable for ordinary people, after all, not everyone can afford a high-performance computer.

Although Crabbox is under the foundation established by OpenAI, Steinberg himself ranks first on the contributor list, with far more commits than the second place. In the past month, the frequency of him updating Crabbox is comparable to the frequency at which he updated OpenClaw back then.

However, as of press time, Crabbox only has more than 1,300 stars on GitHub, which is less than the daily growth of OpenClaw's stars when its popularity just started to decline. It is clear that the market has not responded to it.

The reason why this product did not become popular in China is that domestic AI giants have taken a path more suitable for ordinary consumers. Compared with Crabbox, domestic giants choose to bind cloud execution into their own Agents, which is out-of-the-box, with data never leaving the domain, and does not require additional download, installation and setup like Crabbox.

For example, Moon & Dark's Kimi Claw directly deploys OpenClaw on its own cloud, with pure cloud hosting and no need for local computing power at all. ByteDance's ArkClaw follows the SaaS route, and its enterprise version even uses a confidential computing cluster, where model requests are completed and encrypted in an isolated confidential environment. Tencent's WorkBuddy and Alibaba's DingTalk office services also run entirely on their own cloud servers.

Even so, Steinberg is still maintaining Crabbox with all his effort. Take several updates in August as an example, Steinberg updated it many times late at night local time.

For example, the update on August 9 was at 12:05 local time, the 24th at 3:17 am, the 25th at 1:25 am, and the 26th at 4:36 am.

Steinberg loves development so much, even after finishing a full day's work.

After the Tide Recedes

Although Steinberg's own product did not achieve great success, Codex has achieved great improvement under the influence of his methodology.

For example, Codex introduced the mechanism of multi-Agent parallelism + long task loops, which is exactly the methodology Steinberg himself proposed in OpenClaw.

Under the influence of this set of methodology, the Agent becomes a command center, which can manage several Agents with different divisions of labor to work like managing a team, and the task can last for dozens of hours.

On April 16, Codex added a background computer use mechanism, allowing Agents to open applications and operate interfaces just like humans, and also added an in-app browser. The memory preview launched on the same day also brought in OpenClaw's cross-session MEMORY.md memory mechanism.

In July, Codex was integrated into the ChatGPT desktop application, and computer use became a built-in capability of Codex.

What's more, precisely because it learned from various lessons of OpenClaw including ClawHavoc, Codex launched the Codex Security mechanism, which reviews tasks before execution and predicts what risks the task will cause. Once the risk goes beyond Codex's control, the task will stop immediately.