Zhipu's ZCode Agent Gives GLM5.2 a Frontend Boost That Claude Code Couldn't Match

I've been testing GLM5.2 for two days and was ready to start trashing it.

But I overlooked one important variable: ZCode!

There was a problem in Claude Code that never produced a result. After switching to ZCode, it actually worked, and the visual effect was stunning!

So, I decided to write about ZCode first, and then redo all the tests.

Then this morning I saw that ZCode offers 5 days free for new users and 150% quota for subscribers. I think everyone should just play around with it first!

I had been wondering why Zhipu didn't release an agent-based soft dev kit similar to Codex or Claude, until someone told me they already had one.

When I opened the official website, I thought it was my "Codex" interface. There was even a default project called Gomoku on it!

Actually, you can't say it's exactly the same, but it really looks alike, especially that floating panel on the right!!

Among this type of agent programs, Codex is quite famous and has a good user experience. So when similar software appears, it always reminds people of Codex.

But... their imitation level is still a bit behind mine~~ haha!

I already made a Claude version for them:

It looks like Opus4.8, but actually the built-in model is GLM5.1!

Okay, I was joking when I said they're no good, but I really do have this software!

I originally wanted to share the smooth experience of integrating GLM into JClaude, but it turns out the official version already exists, and it's already at version 3.0.0.

Plans always fail to keep up with changes; some drafts become useless as you write them.

I vaguely remember installing something with a similar name a long time ago, but back then it seemed to be a shell wrapping Claude Code!

Today, I'm going to fully experience this "new software."

Official website: https://zcode.z.ai/

After opening it, the first screen looks like this:

The whole page is very simple! Apart from the header, it's just a software interface, with a little description at the bottom!

It currently supports both Windows and macOS platforms.

After downloading the software, I found the packaging to be extremely cool!

Installation is also very simple, just double-click and next:

After installation, you need to log in:

There are three login methods: one uses the international account z.ai, one uses the domestic BigModel, and one uses an API. I'm a domestic subscriber, so I chose BigModel! If you don't have an account, you can register one; domestic users should prioritize the domestic version.

After logging in, you arrive at the main interface and are immediately greeted with an update:

This interface is very simple and easy to understand!

I noticed it even has phone access:

That's something!

But that's not the point. The point is to test the software's actual capabilities.

So we need to quickly run an example to test its depth.

I've already completed most of my GLM5.2 tests on CC. But one example kept failing:

At first it was stuck for a long time without moving, then retry, and later it just gave an API Error.

This problem was the same in the previous 5.1 version. At that time, I tested continuously for a week, and it was all stuck, rate-limited, not moving even after an hour.

The prompt for this example was:

Design a webpage: Please instantly transform into a retro text adventure game engine. When the user inputs "start", you need to generate a random theme (such as "Survival on a Mars Colony" or "Ancient Cultivation").
​
**Requirements:**
​
1. For each step's choice, generate a real-time **ASCII art illustration** (a picture made of characters) to render the scene atmosphere, without repetition.
2. The game must include a hidden "butterfly effect" logic. If the user chooses "pick up the stone" at step 3, the consequence of this choice must be reflected when encountering a monster at step 10.
3. If the user inputs an unreasonable command (such as "I punch the earth and blow it up"), you need to humorously refuse and guide them back to the plot, without throwing an error.

An infinite-stream text adventure game, integrating butterfly effect design! Very brain-burning!

Since the official version is here, let me test it with this example:

Using this software is very simple. First create a folder as the project directory, then select this folder.

Then set the access permissions in the bottom left corner, confirm the model in the bottom right is GLM5.2, with thinking set to max!

Then click send, and just sit back and wait for the result.

I waited bitterly for 10 minutes without any response, and felt it was hopeless. Until I waited 17 minutes, and it actually produced a result!

The official version really is different! Although it also took a long time, at least it produced a result.

Running isn't the key point. The key point is that this result is super cool!

This is an image here, so you can't see the animation. It has a loading interaction effect, and that interaction effect is one of the best I've seen since testing this example!

In my impression, GLM's frontend has always been lacking.

But this time the test result's interface and interaction are both very good!

Why does GLM5.2 seem like a completely different model inside ZCode?

Is this a lucky draw, or an inevitable result of capability?

Must test more.

Next, let's test it with the "Cyberpunk Version of Along the River During the Qingming Festival."

This example is very strange, so it's rarely been pre-optimized for training by models. It's a very capability-testing example.

This example ran for about 10 minutes, modifying 506 lines of files.

The final result is as follows:

Wow, this completion level is very high! Whether it's layout, color scheme, or element design, it's definitely above average!

Let me give everyone a comparison image:

This was generated in Claude Code.

If we exclude the probability of a lucky draw, the difference between these two effects is still very large.

Damn, does ZCode have some kind of magic?

Why are the UI effects coming out of it so good?

It feels like a completely different model!

I will update all test results to TOPAI today!

https://topai.tonyhub.xyz/

The interface (UI) can be seen at a glance, but the interaction (UX) can only be seen on the webpage. And interaction is something that really tests capability.

This article is mainly to take everyone through experiencing the ZCode software!

I won't expand on a lot of content; let's just start using it.

Next, I will move all test examples here and retest them all.

To confirm two things: one is: Does the UI design really have a bonus? The other is: Does the agent capability have a bonus?

Just getting started, in the "honeymoon phase," the experience is quite good. Next is entering the "trial phase"!

My testing can be very "brutal"~~