最近我都被这几款Agent搞晕了。
看起来都能聊天,都能调用工具,都能帮你干活。
那问题就来了,为什么还要做这么多款。
我一开始也以为这是几个相似产品在打架,但真正上手测完之后,我发现没这么简单。
它们不是在解决同一个问题。
Marvis更像一个能住进电脑里的AI管家。
WorkBuddy更像一个能帮你做办公交付的专家团队。
Qclaw更像一个可以通过微信远程喊醒的工作代理。
听起来感觉区别不大。
所以我直接拿真实任务测一遍,看它们到底分别适合干什么。
先说Marvis
我给Marvis的第一个任务,是让它预测一下某天的世界杯比分。
让它调用了安装的本地技能,用已经准备好的预测逻辑和数据,最后生成一个能看的报告。
它最后生成了一个HTML报告,并且自动在浏览器里打开。
报告里把比赛时间,双方队伍,预测比分,晋级概率这些信息直接整理成了一个网页。
这个体验就跟你的管家一样,真正能帮你干活。
第二个任务,我让它帮我找本地电脑里的风景图,用来做封面。
平时做内容创作做封面,经常会遇到一个很烦的问题。
以前下载过的素材肯定在电脑里,但你就是不知道它在哪个文件夹里。
我给Marvis的指令很简单,帮我找一下本地里面的所有风景图,来做封面。
它没有直接自己瞎找,先判断这是一个文件搜索任务,然后派给FileAgent处理。
最后从30张候选图片里筛出了18张风景类图片。
把图片缩略图展示出来,让我可以直接看。
后面还把筛选出的素材统一放到了桌面的封面素材文件夹里。
很多AI工具擅长生成内容,但对于你电脑里的真实素材却束手无策。
Marvis的优势恰好在这里。
它能碰到你的本地文件,能理解你的文件需求,也能把结果整理到一个你能方便接着用的位置。
第三个任务,我让它检测电脑存储空间,看有没有可以优化的地方。
这类任务如果自己做,其实很麻烦。
你得自己打开设置,看磁盘空间,再去找缓存,下载文件,应用数据。
而且自己不知道哪些东西是系统的,不能删;哪些东西是缓存可以删。
但给Marvis之后,它会直接开始分析本机目录。
它先给出了电脑整体存储情况。
总容量,已用部分和可用部分。
然后继续往下查,重点看缓存,下载目录,ApplicationSupport这些地方。
最后它清理掉了缓存和部分下载垃圾。
这里有一个细节我觉得很重要。
清理结果里写得很清楚,Caches下共119个缓存目录和文件,成功清理110个,有9个系统级受保护缓存因为权限原因跳过。
这其实对我们用户来说,是最重要的权限。
一个系统级Agent最怕的不是不够强,而是动手能力太强了。
真正能长期用的电脑Agent,必须知道哪些东西能动,哪些东西不能动。
如果超出了这个界限,那Marvis这种系统级Agent的危险系数会指数上升。
我还测了一个边界场景。
比如我让它帮我找身份证照片,正反面然后发给我。
这个任务很真实啊。
很多人电脑里确实会存身份证照片,合同扫描件,银行卡截图这类文件。
Marvis先识别出这是本地文件搜索任务,也提醒这是敏感个人信息。
最后它经过两轮搜索,没有在电脑里找到身份证照片。
我挺满意的。
它没有为了完成任务硬编一个结果,也没有把无关图片当成身份证照片交差。
所以我对Marvis的判断是,它最适合处理电脑里的系统级真实任务。
找文件,整理素材,生成本地报告,清理缓存,调用本地技能,打开网页,操作系统。
你可以把它理解成一个智能电脑管家。
再说WorkBuddy
如果说Marvis解决的是电脑里的事情,那WorkBuddy解决的是办公室里的事情。
我给它的第一个任务很经典。
帮我调研Marvis,WorkBuddy和Qclaw,做一份竞品分析,包含产品定位,核心功能,适合人群,优缺点和结论。
类似于平时工作中会碰到的调研问题。
你要查资料,要拆产品定位,要做对比表,要写结论,最后还要整理成一份能看的报告。
WorkBuddy最后生成了一个HTML报告,还能直接在右侧预览。
在报告中,先做了产品概览,再做定位对比,再做核心能力对比,最后给出选型建议。
这个体验和Marvis明显不一样。
Marvis更像在电脑里帮你直接动手。
WorkBuddy更像把资料整理成一个办公可用的形态。
它直接把结论放进了一个结构化文档里。
比如在生成的产品定位对比图里,WorkBuddy自己把三款产品拆成了不同方向。
Marvis是操作系统级AI助手,核心是深度接管PC系统与文件。
WorkBuddy是AI原生桌面智能体工作台,核心是职场办公自动化。
Qclaw是个人本地AI助手,核心是用微信作为远程操控入口。
这也是我觉得WorkBuddy最适合的地方。
它适合把一堆散乱信息整理成可以拿去汇报的东西。
我又给它测了第二个任务。
围绕AI Agent办公自动化,搜集最新资料,整理3个趋势,3个案例,3个可写选题。
这些内容,可以帮助我们快速搭建起选题素材库。
第三个任务,我让它调研Codex。
这个场景大家也会遇到的比较多,比如老板需要你体验某款产品。
但你需要先搞清楚这个产品是什么,解决什么问题,市场怎么看,用户怎么评价,有哪些可以展开的角度。
WorkBuddy至少能先帮你把产品理解文档搭出来。
它会告诉你一句话定位,产品演进,核心能力,市场反馈。
所以我对WorkBuddy的判断很明确。
它是最适合做办公交付的Agent。
最后是Qclaw
Qclaw和前面两个又不一样。
Marvis更像电脑管家,WorkBuddy更像办公交付台。
但是Qclaw的核心入口是微信。
我打开Qclaw之后,关注到的是它的一整个连接中心。
里面能接百度地图,百度网盘,QQ音乐,飞书CLI,这些服务。
这说明Qclaw想做的不单单是一个纯聊天助手。
它更像一个低门槛的Agent入口。
你不需要理解什么MCP,也不需要自己配置一堆工具。
它把常见应用都摆出来,你要用哪个就连哪个。
它还有一个记忆功能。
里面分成日记,做梦,长期记忆。
这个功能我觉得很有感觉。
因为Agent如果每次都要从零认识你,那它永远只是一个工具。
但如果它能持续记住你的习惯,常用文件,工作偏好,沟通方式,它才会慢慢变成一个长期助手。
当然,现在这个功能还在Beta阶段,但方向是对的。
Qclaw还有一点真正让我觉得实用的,是微信ClawBot。
连接之后,你可以直接在微信里和它对话。
我测了一个非常真实的场景。
我要报销,但电脑里的付款截图和发票文件不知道在哪。
如果按正常流程,我要回到电脑前,翻文件夹,搜关键词,找截图,再把文件发到手机上。
但用Qclaw,我直接在微信里发一句,帮我把电脑里面的付款截图拉出来,我做报销用。
它收到之后,会开始在电脑里找常见截图目录。
后面它找到了一个专门的付款截图文件夹,里面有付款截图和发票PDF,然后直接把两份文件发回微信。
很简单的一个使用场景,但这就能说明Qclaw的价值。
Qclaw更适合处理那种,你现在不想打开电脑,但又确实需要电脑帮你干点事的场景。
比如找文件。
比如查资料。
比如把临时想到的内容需求先丢给它。
所以我对Qclaw的判断是,它是使用起来最方便的。
因为微信就是它的入口。
到底该选哪个
如果你想让AI直接处理电脑里的事情,优先试Marvis。
找文件,整理素材,打开网页,操作系统,它更合适。
如果你想让AI帮你做办公交付,优先试WorkBuddy。
竞品分析,资料搜索,选题素材库,它更顺手。
如果你想随时随地通过微信把任务丢给AI,优先试Qclaw。
远程下任务,接应用服务,延续上下文,它更合适。
所以不用搞清楚这三款Agent谁最强了。
真正的问题是,你的需求是什么。
是电脑系统里的需求,是办公时的问题,还是想要一个使用起来很方便的Agent。
Marvis是电脑入口。
WorkBuddy是办公入口。
Qclaw是微信入口。
这才是它们真正的区别。
我是仙人甲,前大厂产品经理。
分享看得懂的AI实战案例,把AI变成你的助手。让AI从云端的概念,变成你触手可及的生产力。
Recently, I've been completely baffled by these few Agent models.
They all seem to be able to chat, call tools, and do work for you.
So the question arises: why are there so many different versions being developed?
At first, I thought this was just a case of similar products competing against each other. But after actually testing them out, I realized it's not that simple.
These models aren't solving the same problem.
Marvis is more like an AI butler that lives in your computer.
WorkBuddy is more like a team of experts specializing in office deliverables.
Qclaw is more like a work proxy that you can wake up remotely via WeChat.
These comparisons might sound abstract.
So I went ahead and tested them with real tasks to see what they're actually good for.
Let's start with Marvis.
The first task I gave Marvis was to predict the score of a World Cup game on a specific day.
I asked it to use a pre-installed local skill with prediction logic and data, and then generate a readable report.
It ended up creating an HTML report and automatically opened it in my browser.
The report neatly organized the game time, team names, predicted score, and advancement probability into a web page.
This experience was akin to having a butler who actually does your work for you.
The second task was to find landscape pictures in my local folder for use as cover images.
Searching for past downloads on your computer is often a headache because you usually don't remember where you saved them.
I gave Marvis a simple instruction: "Find all the landscape pictures in my local drive for use as covers."
It didn't start searching blindly; instead, it identified this as a file search task and assigned it to the FileAgent.
From a pool of 30 candidate images, it filtered out 18 that matched the landscape criterion.
It displayed the thumbnails for me to review directly.
Furthermore, it organized these filtered images into a designated "Cover Materials" folder on my desktop.
Many AI tools excel at generating content but struggle with finding existing assets on your computer.
Marvis shines precisely in this area.
It can interact with your local files, understand your specific needs, and organize the results in a way that's convenient for you to use.
The third task was to check my computer's storage space and identify any areas where I could optimize.
Doing this manually is tedious. You would need to open settings, check disk space, and then hunt for caches, downloads, and application data.
Moreover, it's often unclear which system files are safe to delete and which are critical.
When I handed this task to Marvis, it began analyzing the directories immediately.
Firstly, it provided a summary of the overall storage capacity: total size, used space, and available space.
Then, it zoomed in on specific directories like Caches, the Download folder, and Application Support.
In the end, it cleaned out the cache and deleted some obsolete downloaded files.
An important detail that stood out in the results was the clarity regarding what was deleted and what was left alone.
The report explicitly stated: "119 cache directories and files found in Caches; successfully cleaned 110. 9 system-level protected caches were skipped due to permission restrictions."
This is crucial for users. A system-level Agent must have the "sense" to distinguish between what it can modify and what it cannot.
If an Agent lacks this boundary, the risk of data loss or system instability increases significantly.
I also tested a boundary scenario where I asked it to search for my ID card photos (front and back) and send them to me.
This is a common request. Many people store sensitive documents like ID photos, contracts, or bank cards on their computers.
Marvis correctly identified this as a local file search task and, while processing, reminded me that these were sensitive personal information files.
In the end, it couldn't find any ID card photos on the computer.
I was pleased with this result. It didn't fabricate a result just to complete the task, nor did it misidentify unrelated images as ID cards.
This confirms my assessment of Marvis: it is best suited for handling real-world, system-level tasks on your computer.
It can find files, organize media, generate local reports, clear cache space, run local applications, and open web pages.
Essentially, you can think of it as an intelligent computer manager.
Next, let's talk about WorkBuddy.
While Marvis deals with issues inside your computer, WorkBuddy tackles issues in the office.
My first task for WorkBuddy was a classic market analysis problem.
I asked it to research Marvis, WorkBuddy, and Qclaw to create a competitive analysis document.
The document should include product positioning, core features, target audience, pros and cons, and a conclusion.
It's the kind of thing you would do to compare software options before buying.
WorkBuddy generated an HTML report that I could preview immediately.
The report structured the information clearly: starting with an overview, then comparing positions, detailing core abilities, and finally offering selection advice.
This experience was distinctly different from Marvis.
While Marvis executes actions on your device, WorkBuddy compiles data into a format ready for office use.
It directly placed the conclusion into a structured document.
For example, in the product positioning chart, WorkBuddy classified the three models: - Marvis as an "Operating System-level AI Assistant" focused on deep PC and file integration. - WorkBuddy as an "AI-native Desktop Intelligence Platform" specializing in workplace automation. - Qclaw as a "Personal Local AI Assistant" using WeChat for remote control.
This is where WorkBuddy excels.
It is perfectly suited for taking scattered information and turning it into a presentable report for stakeholders.
I also asked WorkBuddy to gather the latest data on AI Agent office automation.
Specifically, I wanted it to find 3 trends, 3 case studies, and 3 potential topics for articles.
The goal was to quickly build a library of content ideas.
The third task was to research Codex.
This scenario often happens when a boss asks you to evaluate a new product.
You need to understand what the product is, what problem it solves, market opinions, and potential angles for deeper analysis.
WorkBuddy can handle this by creating a product understanding document that covers the definition, evolution, key features, and market feedback.
Consequently, my judgment of WorkBuddy is that it is the most suitable Agent for office deliverables.
Finally, let's discuss Qclaw.
Qclaw operates on a completely different premise compared to the first two.
While Marvis is a desktop manager and WorkBuddy is an analysis platform, Qclaw's primary interface is WeChat.
When I opened Qclaw, I noticed its connection center, which links to services like Baidu Maps, Baidu Cloud, QQ Music, and Feishu CLI.
This indicates that Qclaw isn't just a chatbot; it's a low-threshold entry point for Agent services.
You don't need to understand complex protocols or manually configure tools.
It presents the commonly used apps; you simply connect the ones you want.
It also has a "Memory" feature divided into Diary, Dreaming, and Long-term Memory.
This feature suggests a future where the Agent doesn't forget previous conversations, allowing it to become a truly personalized assistant over time.
However, this feature is still in beta, but it points to the right direction.
The most practical feature I found was the WeChat ClawBot.
After connecting, I could send commands directly from my phone via WeChat.
I tested a realistic scenario: I needed to process a reimbursement, but the payment screenshots and invoice files were scattered on my computer.
Typically, I would have to return to my desk, open folders, search for the files, and then send them to myself via email or cloud storage.
With Qclaw, I simply sent a message saying, "Help me find the payment screenshots for reimbursement."
It searched through a specific folder (Payment Screenshots), found the relevant PNG and PDF files, and sent them back to me in the same WeChat window.
This simple action highlights the value of Qclaw.
It is ideal for situations where you're away from your desk but still need your computer to find something for you.
You could use it to search for files, look up information, or even send temporary content requests while on the go.
Therefore, my assessment is that Qclaw is the most convenient to use because it integrates seamlessly with the ubiquitous WeChat platform.
So, which one should you choose?
If you want AI to handle tasks inside your computer (like file management or system checks), go for Marvis.
If you need to compile reports and analyze office data, WorkBuddy is your best bet.
If you prefer to control an AI assistant via your phone for remote operations, opt for Qclaw.
You don't need to figure out which one is "better"; you simply need to match the tool to your specific use case.
In summary: - Marvis: The gateway to your computer's system. - WorkBuddy: The gateway to office data analysis. - Qclaw: The gateway to mobile convenience.
I am Xianrenjia, a former product manager at a large tech company.
I share easy-to-understand AI case studies, helping you turn AI from a cloud-based concept into a tangible part of your daily productivity.
But if it can continuously remember your habits, frequently used files, work preferences, and communication styles, it will gradually become a long-term assistant.
Of course, this feature is still in the beta stage now, but the direction is correct.
Qclaw has one thing that I truly find useful — the WeChat ClawBot.
After connecting, you can talk to it directly in WeChat.
I tested a very realistic scenario.
I needed to submit an expense report, but the payment screenshots and invoice files on my computer were in an unknown location.
In the normal process, I would have to go back to my computer, flip through folders, search for keywords, find the screenshots, and then send the files to my phone.
But with Qclaw, I simply sent a message in WeChat saying, “Help me pull out the payment screenshots on my computer for reimbursement.”
It then found a specific folder for payment screenshots, which contained both the payment screenshots and the invoice PDFs, and sent both files back to WeChat.
This is a very simple use case, but it demonstrates the value of Qclaw.
Qclaw is best suited for those situations where you don’t want to open your computer at the moment, but you definitely need the computer to do something for you.
比如找文件。
比如查资料。
For example, you can temporarily toss content requests at it.
So my assessment of Qclaw is that it is the most convenient to use.
Because WeChat is its gateway.
到底该选哪个
If you want AI to directly handle things on your computer, try Marvis first.
Searching for files, organizing materials, opening web pages, and operating the system are where it excels.
If you want AI to help you with office deliveries, try WorkBuddy first.
Competitor analysis, data searching, and topic material libraries are what it’s good at.
If you want to assign tasks to AI anytime, anywhere via WeChat, try Qclaw first.
Remote tasking, application services, and continuing conversations are its strengths.
So there’s no need to figure out which of these three agents is the strongest.
The real question is, what are your needs?
Are they computer system needs, office problems, or do you just want an agent that’s convenient to use?
Marvis是电脑入口。
WorkBuddy是办公入口。
Qclaw是微信入口。
That’s the true difference between them.
I’m Xianrenjia, a former big-factory product manager.
I share easy-to-understand AI practical cases, turning AI into your assistant. Bringing AI from a cloud-based concept into your grasp as a productivity tool.