A Deep Dive into Zhipu AI's Full Stack Sandbox Coder

The GLM 4.5 model is a pretty good model. There's also a smaller model, the Air variant, which is something you can run locally on a high-tier MacBook, which is quite awesome and pretty capable as well. It is available for a pretty cheap API pricing; it costs very little and it's great at tool calling and everything like that. Plus, you can also use it completely free without any limits on their chat platform. It's quite good for sure.

Introducing the "Full Stack" AI Coder

However, there's also something that was launched alongside that, and that is a new option called Full Stack. I found this option pretty amazing and thought I should talk about it.

To get started, you'll have to come to their ZAI chat platform and then just get signed up. Once that is done, you'll see the chat interface. You can select the model here and use it as much as you want. But, as you can see, you also have the Full Stack option. This is a pretty cool one.

What this does is pretty amazing. It basically spins up a sandbox, and then the coder here codes in that sandbox and creates whole projects with databases and multiple files, and it can code in any language as well because it can actually run any kind of terminal command. If you think about it, then it is like an agent but without the option to do web searches or similar stuff. It's just focused on coding. So that is basically an overview of it.

A Practical Example: Building a 3D Game

Let's try to use it as well. For example, I'm going to ask it to build me a car driving three.js game.

Build me a car driving three.js game.

Now, once we send it, you'll see that it will first spin up a virtual sandbox for us. It's pretty snappy and it gets started quite easily. Then it will go ahead and open up a panel on the right as well, where you'll see the files and the preview once the code is generated.

Key Features and Capabilities

You can also see that it can easily run terminal commands here as well. Here it is running terminal commands and checking the files and stuff. You can see that it will run a bunch of commands in order to check what code is already there and similar things. It will also use the terminal to install packages that are not available or needed in order to make sure that it all works.

You can see the files and other stuff in the workspace by hitting the option here. And then you'll see that it will show you all the files in the workspace. It also creates a to-do list here as well, and it is really good at planning and stuff like that. It is pretty detail-focused. It's amazingly good for creating full-stack apps because it rolls in a loop until the stuff is finished, and it can test and everything like that as well. It also thinks as it does the edits, and after each command or edit, it thinks again, which is pretty awesome.

Anyway, you can see the diff of the changes that it has done for you here. It is really good at coding and stuff with this agentic loop. It's really well done. In a bit, it gets done, and you can see the preview on the right, and it looks pretty awesome and works well. It is amazingly detail-focused. It's like other agents but extremely focused on coding instead. And it is pretty awesome. If they add web search capabilities to it and maybe browser capabilities, then it can become an awesome alternative to stuff like that, which is great.

One thing is that it can be a bit unstable at times because their infrastructure is overloaded at this time, and it can give you some errors at times or something. But just refresh and it should work fine again.

Project Management and Deployment

You can also download the whole project and workspace as well with the download option. It will give you a zip file that you can download and then use the source code locally. It's pretty awesome if you ask me, as you can continue on the project locally quite easily.

You can also share the preview with others as well by clicking this option, which will open up the preview here, and then you can share it with others and let others use it too. Also, if there are multiple routes, then you can also go to them through the option here at the top as well.

Another option is Publish. Publish will basically publish it to a URL, and that would mostly run 24/7, which is also great. If you create a small app and want to let it run, then that can be done. So, this is quite good. It is a really awesome coder that actually gets a sandbox and then works on its own. It is capable for most stuff, but if it had internet surfing and browser capabilities, then it would have been even better for sure. I think that they can add it soon enough.

Limitations and Considerations

Now, there are some limitations, and one of the major ones is that it only allows you to have a couple of workspaces per account at once. So if you want to work on a third project, then you'd have to delete one of them before proceeding. Apart from this, there is not much limit, which is pretty good.

This is an amazing model for sure. Also, if you get more errors because of the high traffic, then you can switch to the Air model, and that has no errors at all. That is also pretty good and works with the full-stack stuff as well. It is also faster, and you'll get the responses super fast with that.

That is majorly how it works. I found this pretty amazing, and you can go ahead and use this as well and see what sticks for you and what doesn't. I find this really amazing because there are several coders that perform similarly, but this is a new take on that thing, and it works pretty well. You can go ahead and use it as well and see how well it performs for you. Overall, it's pretty cool.