Review of ten AI lobsters, who is the top domestic lobster?
(The author of this article is Light Cone Intelligence, published by Titanium Media with authorization)
Text by Light Cone Intelligence, authored by Wei Linhua, edited by Liu Junhong and Wang Yisu
OpenClaw's explosive popularity at the start of the year has kept us feasting on 'domestic lobsters' for three whole months.
From the cloud versions served first to the later local small lobsters claiming to offer a 'native lobster experience,' not only were users overwhelmed, but even we, who had to chase product reviews daily, felt dizzy.
What they ignited was not just a product, but also an imagination — 'letting AI work for me.'
As big names flaunt the massive Tokens consumed running lobsters and social media floods with cool screenshots of 'letting AI do my work,' countless office workers have sparked a simple desire: I want a small lobster that can work for me. Ideally cheap, user-friendly, and more reliable than my colleagues.
But the excitement belongs to the manufacturers, and as a user, I feel a sense of emptiness: I installed it on my computer as soon as it was released, but the error messages made my head spin; I might not be able to finish two tasks in a day. Complex tasks seem beyond its reach, so why shouldn't I use Manus instead?
Among all these lobsters, which one can give me a painless experience comparable to OpenClaw?
With this in mind, Guangzhui Intelligent reviewed 10 small lobster products on the market from the perspective of a user without any AI background, to see if they can withstand rigorous testing.
Given that some users simply wish to try out using 'shrimp' for basic tasks, while others aim to become 'shrimp catchers' and evolve together with shrimp towards silicon-based life forms, we've tailored our reviews to different needs, progressing from simple to complex: starting with basic tasks like scheduled daily reports and information gathering, then advancing to see if these shrimps can master Skills and guide me step-by-step through complex tasks performed by experts.
To start with the conclusion, most shrimps can handle simple tasks. But for more challenging work, the majority turn into 'time killers,' and success isn't guaranteed.
Who can make users become 'shrimp hunters,' and who turns them into 'shrimp slaves'? We conducted a comprehensive review.
At the very beginning of 'eating shrimp,' I was in a great mood because the installation experience for every product was incredibly smooth.
If you've ever tried deploying OpenClaw on your own without development experience, I bet you've wasted more than a day of your life, otherwise, there wouldn’t be a business offering on-site OpenClaw installations for thousands of dollars.
The contribution of domestically produced crayfish is bringing the threshold of 'crayfish' down from professional-grade to consumer-grade:
Among these, cloud-based crayfish can generally be used out-of-the-box without any user intervention; interacting with cloud-based shrimp is as easy as opening an AI model's dialogue box. Installing local crayfish isn't difficult either—it’s similar to downloading normal PC applications, as long as you know how to download the installer from the official website, there’s no major issue.
Installation is just the starting line; from configuration onwards, it’s time for vendors to showcase their unique strengths.
If you don’t want a cold AI assistant and prefer it to feel more human, that’s easy—you can set the desired personality for the shrimp.
For instance, Feishu, Stepwise, and Baidu's Lobster all come with currently trendy personality setups (Soul.md), allowing you to define how the lobster should address you and use prompts to describe the kind of 'personality' you want your lobster to have, making its interactions feel more like real human conversations.
The personality I configured for my lobster on DuClaw
I turned this batch of lobsters into a 'reliable but love-to-complain colleague.' As a result, Stepwise Lobster would grumble about how complicated procedures are when running tasks, while Baidu would say, 'Leave it to me.' The coldness typically associated with AI disappeared, and these cybernetic colleagues with a bit of attitude seemed less annoying even when they encountered errors.
If you can only use AI in front of a computer, its convenience is greatly diminished. The original intention behind creating the 'Lobster Father' was to find an assistant for remote work, so whether or not it can be accessed via mobile devices is also a key feature.
Compared to OpenClaw, which requires painstaking manual configuration,Major domestic IM platforms have started actively providing backdoor access to lobsters; now most only require users to scan a QR code and wait a few minutes before the platform automatically configures everything for you.
For example, WeChat has developed plugins that welcome various lobsters to connect via scanning codes, while products like Feishu and QQ can now establish connections with just one scan.
Once the lobster is set up and capable of sending messages to your phone, we can officially start putting it to work.
When it comes to actual work, the gap between imagination and reality becomes apparent: the joys and sorrows of users do not resonate, and the intelligence of lobsters varies significantly.
Take testing the AI daily report task as an example. This is a scheduled task requiring the AI not only to scrape the necessary information from various sources according to your requirements and compile it into a daily report but also to send it to you at a fixed time every day.
The test results were surprising: based on the standard of 'getting it right the first time,' we immediately eliminated half of the products.
Among them, those that managed to deliver on time at the first attempt included Zhipu, KimiClaw, MiniMax, QClaw. The remaining ones encountered various error reasons and still required manual intervention to 'fix the assignments.'
The difference between cloud and local is particularly evident here.For users who cannot afford dedicated equipment (such as a Mac mini), if the local lobster system shuts down or loses internet connection, the execution of scheduled tasks could be affected. In contrast, the cloud version can provide stable daily delivery without being restricted by the status of the local device.
When evaluating from the perspective of content quality, AutoClaw from Zhipu, Claw from AliCloud, and Duclaw from Baidu provided richer and more comprehensive information, ensuring that the content was fresh from the previous day. However, some made timing and factual errors. For example, KimiClaw mistakenly treated last year’s news as this year's, with obvious mistakes.
A lobster capable of producing daily reports can only be considered an average performer. Workers also need AI to handle simple tasks in their jobs to see if it can truly manage various chores.
Using the higher-demand 'text-to-image' task as the testing standard, we asked each lobster to create a previously viral Nano Banana-style cartoon infographic titled 'An Introduction to xxx.' The subject of introduction was themselves.
In terms of the final output quality, JVS Claw from Alibaba stood out significantly. It sourced a personal user-uploaded skill from Vercel’s official Skill website and generated five product introduction images at once. Although it utilized a Xiaohongshu image generation Skill, the overall style met the requirements for a cartoon explanation.
Apart from Alibaba, Step Stars also utilized a skill from its own aquatic market, explicitly named Nano Banana. Though the resulting image was in English, the cartoon style was achieved, fulfilling the one-image explanation requirement.
Other products also generated images either by providing text-to-image prompts or through API integration. While they all produced results, the styles were vastly different from what I desired.
Not buddy, you give me such a picture when generating an introduction?
In short, the effectiveness of task execution ultimately depends on the comprehension ability of the model integrated into Xiaolongxia and whether the Skill library is rich enough. Although all are connected to the Gemini drawing model, the effect of the generated images still varies greatly due to Xiaolongxia’s understanding and invocation situation.
There's often a huge gap between 'usable' and 'user-friendly'.
The core of advanced usage lies in the Skill ecosystem.
Why are the Xiaolongxia used by internet experts so powerful? Today it can be Jarvis, tomorrow it can act as a financial manager? To unlock imagination and enable Xiaolongxia to handle more complex tasks, users do not have the patience to type hundreds of words teaching AI how to work.
The rich Skill ecosystem on the internet equips Xiaolongxia with customizable 'claws'. The Skills grown within the open-source ecosystem come from contributions made by every developer—
When there is a long-term demand for repetitive tasks, such as checking emails every day to confirm schedules, this set of prompts given to the AI can be standardized. Next time, this Skill can be directly selected for execution. Teaching children might not guarantee a 100% return, but teaching Xiaolongxia does.
The quantity and quality of Skills represent the extensibility of Xiaolongxia.
Pre-installation by manufacturers is the beginning of a good user experience. I let Xiaolongxia search for the number of pre-installed Skills in these products and make a table for me. Zhipu stood out by finding all products completely and providing mostly correct results.
The table provided by Zhipu AutoClaw
Egregiously wrong are Tencent's QClaw and MiniMax's MaxClaw, which failed to comprehend the instruction 'products对标OpenClaw,' mistaking them for Agent products like ByteDance’s扣子. Baidu didn’t even filter out any relevant product, instead targeting entire companies in their statistics.
Among these, three types of Skills have become essential for assembly:
Creator allows users to create their own Skills on demand;
Find Skill eliminates the need for users to visit a Skill website to download and install manually—it automatically finds and installs the necessary Skills in the background. Vetter ensures that the installed Skills are safe, reviewing each one before installation to prevent malicious Skills from harming your computer.
However, some Skills fail to deliver their intended effects even after installation.
For instance, Baidu’s Duclaw also includes security-related review Skills, but its approach is to install first and then warn users about potential risks. Only after we pointed this out did they indicate they would 'review beforehand next time.' This 'next time' seems to come too late.
The quality of the Skill ecosystem is also crucial
With several Skill websites already established overseas, many domestic products have opted to build their own Skill ecosystems. Currently, official Skill stores run by Tencent, StepStellar, and Cheetah include such preparations. For example, StepStellar has built an aquatic market with over 5,000 Skills, encompassing both official and user-uploaded ones. The Nano Banana-related Skill mentioned earlier was sourced from their self-built 'aquatic market.'
Example: In EasyClaw’s Skill store, the same Skill used in Fu Sheng’s version of crayfish is highlighted.
Skill is indeed important, but can the crayfish find the right Skill based on my needs?
We let these crayfish look for a skill — the popular 'Crayfish Office' visualization project from some time ago. Through this office interface, you can see whether the crayfish are working, thinking, or slacking off in front of the couch. QClaw, which has this feature built-in, skipped this test.
Even though I don't have the energy to exercise after work, the crayfish can still lift weights. Image source: QClaw
I had them help me collect Skills that could build the 'Crayfish Office.' Most were able to find the right project, but their performance varied when it came to execution.
Alibaba's JVS Claw failed to load once before running successfully; EasyClaw installed successfully on the first try, making it one of the quicker ones. Zhipu misunderstood the task and ended up installing a dashboard with no linkage or office interface. Some even tried to write code for me, as Shen Teng would say, 'Some people are both foolish and hardworking.'
As you can see, simply finding and installing based on a description is no longer a challenge for most crayfish. However, many problems arise during the subsequent chain of execution.
We then assigned a slightly more complex task: having the crayfish connect to my email, organize the content of unread emails, so that in the future, I wouldn't need to check my emails — the AI would just tell me what I've received.
Configuring email seems simple at first glance, but upon closer inspection, it’s full of complications. I asked the AI to handle it by connecting through an email API, which also involved teaching me how to set up the relevant configurations, guiding me to activate the email API. During the process, issues like refresh token expiration arose, and the crayfish had to help me figure out how to resolve timing issues.
This is StepClaw's summary of all the work it completed.
What seemed like a simple API connection took over three hours with these dozen crayfish. I wanted the AI to save me time, but the amount of time spent teaching them made me question the whole endeavor.
The first to complete the task was StepClaw from Leap星辰. Although it repeatedly hinted that I could manually import email data for analysis (just like a colleague pushing me to take action), after I insisted on 'not lifting a finger,' it bypassed the hurdle of obtaining a token by writing its own script that could run on the web to retrieve the token itself. Despite its constant urging to 'hurry up,' it eventually connected successfully.
My first successful connection after continuous complaints.
Later, Kimi Claw also provided me with an automatic token retrieval script, but ultimately it failed as the script wouldn't execute. Zhipu's AutoClaw insisted I use the command line, but mostly there was no response. MiniMax gave increasingly abstract links and scripts that I couldn't run, which led to failure. EasyClaw struggled with environment issues, failing twice before attempting solutions, but ultimately no reliable method emerged.
QClaw and Baidu’s DuClaw, as well as Alibaba’s JVS Claw, opted for shortcuts. They utilized simpler methods where I set a dedicated application password in Google, allowing access without retrieving my actual login credentials. Both Alibaba and Baidu succeeded on their first attempt; notably, Baidu remembered my previous request and directly sent over the summarized email results—great job!
QClaw saw but didn’t respond.
However, QClaw seemed blocked by system settings, frequently encountering 'difficulties' and going dormant. It failed to respond four out of six times, let alone address any issues.
It can be said that even when successful execution is possible, those without a programming background can only follow the steps outlined by others, hoping for success through trial and error—either succeeding or losing patience along the way.
Why do these tools show such varied performance? For relatively complex tasks, what is tested is the ability to configure models and design Harnesses.
The former determines whether the model can leverage its Agent-related capabilities to build useful tools and resolve environmental issues. The latter refers to Harness, a term literally meaning horse harness but similarly applied to Agents—Harness acts as the shell around the Agent, encompassing all engineering configurations.
The model’s capability determines whether AI can autonomously find solutions when encountering problems. During testing, we discovered thatThe saying 'you get what you pay for' applies equally in the AI field.
For instance, Zhipu, which I found quite useful, cost me 300 credits for a single statistical table task (Zhipu offers a free quota of 500 credits). In comparison, although QClaw may not be as good, its affordability likely has to do with its generous built-in model, given that it boldly offers me a daily consumption limit of 40 million tokens.
QClaw is generous!
Since most OpenClaw-like products generally don’t support external models, this is particularly evident among large-model startups and cloud providers. However, local products like EasyClaw and QClaw still allow it. Given the differences in models, it's hard to compare Harness functionality.
But when evaluating stability and self-repair capabilities, some products have shown clear issues. For example, both EasyClaw and StepClaw encountered errors while I was using them; the former lacked a 'gateway restart' option to resolve the issue, and while the latter promotes using its own Agent assistant 'StepAgent' to fix StepClaw, my repeated attempts yielded poor results.
By the way, I was completely baffled by those two until Alibaba’s JVS Claw guided me step by step on how to write specific gateway restart commands for the Windows system, which eventually fixed the problem.
Alibaba Shrimp and I painstakingly figured out the command line together.
By now, you’ve probably realized that the potential ceiling unlocked through raising shrimp is actually quite high—it all depends on how you want to use it:
Various Skill websites are like stores filled with secret martial arts manuals: creating viral Xiaohongshu content, or having the shrimp 'self-learn and evolve' every morning—there are countless creative ways to explore. If you want to learn more about imaginative applications across different scenarios, the rest is up to AI to handle for you.
However, the extent to which these can be achieved, whether they’re stable, and whether they can 'learn by analogy,' depends on each product’s unique capabilities powered by their models and Harness.
Like Manus, which was accused of being a copycat and yet had few rivals within a year and wasn’t successfully imitated by major companies, these alternative products to OpenClaw need more effort to truly evolve into something usable. The next step is how to quickly iterate on the product so that users no longer complain about frequent crashes and error-prone 'lobsters'.
After my computer kept popping up with inexplicable command line interfaces, installing dozens of lobsters caused my C drive to overload (because some lobsters don’t support changing the workspace to the D drive), and a series of side effects, the evaluation results have basically come to an end.
In terms of stability and usability, the cloud-based option that stands out is Alibaba Cloud’s JVS Claw. When facing various issues, it rarely reported errors, and its performance in daily report tasks and email configuration was quite satisfactory.
Compared to similar cloud deployment products, it also excels in social features. For instance, both Baidu and ByteDance’s lobsters require uploading images via cloud storage files; ByteDance’s ArkClaw even needs manual configuration of the cloud drive or taking over a cloud computer for uploads. However, Alibaba's version allows direct uploads, similar to KimiClaw and MaxClaw, which are built on Agent-based foundations. Moreover, the cloud computer setup means it can handle simulated local tasks in the cloud, whereas KimiClaw lacks this cloud computer mode.
On the local product front, the standouts are from Step Stars and Zhipu.
Among them, Zhipu’s AutoClaw has superior stability, almost never reporting errors, and delivers top-tier performance in information search and summary table tasks. Step Stars, though less stable and giving off the impression of being a 'task-pushing lobster', still performs admirably in daily reports and email connection tasks. It can create web tools to handle tasks, offering an experience similar to lobsters automatically finding tools and connecting APIs.
The middle-ranking options are KimiClaw, MaxClaw, QClaw, and DuClaw. The first two have no stability issues but perform at a moderate level in tasks. The latter two occasionally encounter unresponsive errors, but no irreparable problems occurred, possibly related to server issues, and their task performance is also average.
The worst performers were WorkBuddy and ArkClaw. Both struggled with continuous use. For example, WorkBuddy experienced large-scale errors twice, although the first time was due to a massive influx of traffic. Subsequently, there was a two-day period of unresponsiveness, but after recovery, response speed improved and reached a passing level. ArkClaw, on the other hand, typically responds only after 2-3 queries. When normal use becomes a luxury, testing specific task performance is out of the question.
Regardless of the form, stability and task success rate are the core metrics determining user experience. No matter how fancy the features are, nothing beats running stably even once.
Of course, the race to determine 'who can replace OpenClaw domestically' has just begun.
Compared to抢先占据市场 through early releases, subsequent updates and maintenance will decide whether these apps can remain on users' computers and phones instead of being uninstalled shortly after an initial trial.
A comparison between cloud-based and local products shows that cloud solutions are clearly better suited for current user demands for computer security. After all, changes made locally to system configurations or files might lead to losses that cannot easily be reversed. However, in terms of functionality expansion, with broader access permissions locally, Little Red Shrimp offers a wider range of tasks and more impressive performance.
As the first wave of reviews winds down, we've seen the release of the Button Edition of Little Red Shrimp and a major update for QClaw V2. Amid user complaints about poor usability and high costs, Little Red Shrimp continues to accelerate its iterative evolution.
The breakout shrimp app may well be just around the corner.
Risk Disclaimer: The above content only represents the author's view. It does not represent any position or investment advice of Futu. Futu makes no representation or warranty.Read more
Comments
to post a comment
1
5