Yunzhisheng's U2Claw: A Desktop AI Agent That Actually Works—Or Just Old Wine in New Bottle?

Let’s be honest—when I saw Yunzhisheng announcing U2Claw, my first thought was: another “intelligent assistant”?

Here’s the thing. Over the past two years, I’ve tested at least twenty AI Agent products, from AutoGPT to Claude Code, from various “digital employees” to “smart butlers.” Most of them look great in demos but fall apart in real use. So when I saw U2Claw branding itself as “desktop-level,” I instinctively raised an eyebrow.

But after digging into the technical docs, I realized this one’s different.

What is it, actually?

U2Claw is essentially a local-first AI Agent gateway. What does that mean? It runs on your own computer, you chat with it through messaging apps (WeChat, DingTalk, Feishu all work), and it can call your local tools and software—browser, Office, dev environment, file system, even operate your mouse and keyboard.

How’s that different from cloud-hosted Agents?

The biggest difference: it can actually touch your local environment.

Cloud Agents can at most help you call a few APIs and process some text. But U2Claw can directly open Excel on your computer, fill in data and create charts per your instructions; open browsers to fill forms or grab tickets; even operate your IDE to write code and run tests.

My personal take: this is what “true desktop-level” means—not the “fake desktop” that just chats on a webpage, but something that can actually touch your local workflow.

Can it actually do the work?

Don’t rush, let’s look at the data.

Yunzhisheng’s website lists several use cases: office automation, browser automation, multi-channel interaction, persistent memory and planning. I tested two of the most practical ones.

First: “Excel auto-processing.”

I gave it a task: “Open the sales_data.xlsx on D drive, summarize by month, and create a bar chart.” In less than two minutes, it actually did it—opened Excel, read data, created a pivot table, inserted a chart, saved the file. The whole process was faster than me teaching an intern step by step.

But there’s a catch: if your Excel file path has Chinese characters, or if the file format is slightly complex, it gets stuck. Tried three times, succeeded twice—not a high success rate, but at least it shows it can do the work.

Second: “Browser automation.”

I asked it to check price comparisons for a product across different platforms. It opened Chrome, visited JD.com, Taobao, and Pinduoduo in sequence, scraped price info, and compiled everything into a table. This workflow was more stable, basically no errors.

Honestly, while these two scenarios aren’t complex, they already cover pain points in many people’s daily work. Compared to those AIs that only know how to “chat chat chat,” U2Claw has at least taken the first step toward “can actually do work.”

Three things worth noting

First, the local-first architecture design.

U2Claw’s core logic: data doesn’t leave your computer, computing power uses your local hardware. This is a plus for privacy-sensitive scenarios (like processing internal company files). Yunzhisheng is also pushing a “sell service + sell computing power” business model, making software free and earning through token billing and API calls—more sustainable than one-time license sales.

Second, flexible model adaptation.

It doesn’t lock you into a specific model—you can connect GPT, Claude, DeepSeek, even locally deployed open-source models. This design is smart, avoiding the risk of “being held hostage by model vendors” and giving users more choice.

Third, persistent memory and planning capabilities.

It can remember your previous task context and automatically associate it next time you chat. For example, if you asked it to track a project’s progress yesterday, and today you ask “how’s that project from last time?” it can directly pull up yesterday’s status and continue updating. This capability is particularly useful in long-cycle tasks.

Of course, there are issues too

Let’s talk performance first.

I ran U2Claw on a MacBook Pro, 8-core M2 chip, 16GB RAM. Response speed was okay, but occasional stutters happened. If your computer specs are average, the experience might take a hit. After all, local inference demands decent hardware.

Then stability.

As I mentioned earlier, Excel processing success rate isn’t high, and slightly complex tasks tend to fail. Yunzhisheng admits they’re still in a “technical iteration period,” with many features not yet mature.

Finally, ecosystem.

U2Claw currently supports limited tools and software, mainly covering office and programming scenarios. If you want it to help with more specialized work (like video editing, 3D modeling), it can’t do that yet.

Is it worth trying?

If you’re like me, handling lots of repetitive work daily (filling forms, organizing files, researching), U2Claw is worth a try. Especially the local-first, privacy-secure features—it’s attractive to many enterprise users.

But if you expect it to be an “omnipotent” AI assistant like in sci-fi movies, you might be disappointed. Its current positioning is more like an “automatable automation tool,” not a “thinking digital employee.”

I personally think Yunzhisheng’s direction is right—from “selling software” to “selling service,” from “cloud hosting” to “local-first.” This represents AI Agent products evolving from “showing off” to “being practical.”

Don’t ask me if it’s truly intelligent—ask yourself if it can save you time. If yes, it’s a good tool; if no, it’s just old wine in a new bottle.

As for which category U2Claw falls into, you’ll know after trying it yourself. I’m planning to keep using it, see how it evolves.