At the beginning of the year in January, I saw news that ByteDance launched an AI IDE for overseas, calledTrae.

I thought, all the current IDEs have AI plugins, complete in function, Byte also has them.MarsCode PluginIs it necessary to create a separate IDE?
I haven't paid special attention to this tool.
Last week, I met a colleague from ByteDance and learned that the domestic version of Trae will be released this week. I also got an early access version, which is the first AI-native IDE for the domestic market.
Today, let's talk about my usage experience and explore a bigger issue at the same time:With so many AI tools already available, should we keep reinventing the wheel?
Let's start with the conclusion: Trae's Chinese version is worth using. As an AI IDE, it's generally better than plugins, and it includes built-in large models—Doubao 1.5 Pro, DeepSeek Full Version—available for unlimited free use.
I. International Version vs. Chinese Version
Before sharing my experience, let me explain the main differences between the two versions, which primarily lie in the models.
The international version of Trae uses foreign models, which can lead to issues like connection failures, long waiting times, and slow internet speeds for Chinese users.
The Chinese version of Trae uses domestic models, ensuring stable and fast connections, with the interface also customized to suit Chinese user habits.
I also asked a ByteDance colleague about the relationship between Trae's Chinese version and MarsCode. The answer is, they are products of the same team but designed for different scenarios.
If you want to keep the originally used IDE and only add AI programming-related features, then use the MarsCode plugin. If you want to experience a completely new native IDE designed around AI, then try Trae.
These two products will be developed in the long term, and you can choose according to your needs.
2. Installation
Trae's domestic version needs to go to the official websitetrae.com.cnDownload.
Currently, there are three versions: Windows version, macOS Intel chip version, macOS M-series chip version.
After installation and startup, the following screen will appear.

Next, choose the light/dark theme and language (Simplified Chinese).

It will also ask if you want to import the original IDE (VS Code/Cursor) configuration, which is quite thoughtful design.

Finally, it asks if you want to log in. You can use it without logging in, but some features will be restricted. If you log in, it will redirect you to the Trae official website to log in with your phone number.

Once everything is done, you will enter the user interface.

III. User Interface
Trae's user interface is quite simple. The left side is the VS Code interface. If you're familiar with it, there's no difficulty in getting started.

The right side of the interface is the AI area. You can see it is divided into two modes: Chat and Builder.

The Chat mode is the default mode, which is the Q&A mode. The AI answers questions and can also generate code, but it won't generate projects. You need to manually copy the code over.
Builder pattern is a project pattern, you provide a text description, let AI generate the project with one click.
Below, let's try these two patterns.
Four, Chat mode
The main body of the Chat mode is a dialog box, the button in the lower right corner can select the underlying model.

The domestic version currently has three models available for selection.
- Bean package 1.5 pro
- DeepSeek R1
- DeepSeek V3
These three models are unlimited free use, all DeepSeek series models are full-power versions。I tested it, the speed is satisfactory, it can provide answers relatively quickly without encountering the "server busy" prompt, even for the most time-consuming DeepSeek R1 model.
As for the quality of the answers, to be honest, the leading large models at this stage are already hard to find fault with in this regard.

You can ask any question, including non-technical ones.

I directly had it generate code, and the response format is very user-friendly.

Everyone注意, in the image above, the first code block is terminal code, with three buttons in the upper right corner (see the image below).

The three buttons in the upper right corner of the image above are "Copy," "Add to Terminal," and "Run." If you click the latter two buttons, it will directly input the code into the built-in terminal in Trae, just like the image below.

This part, I feel, is smoother to use and has a stronger overall feel, which is probably the advantage of making AI into an IDE.
5. Builder Pattern
Next, let's look at the Builder Pattern, which can generate a project with one click.
Click the tab above to switch to this mode. You can see that on the first entry, there will be a "Start Builder" button (see image below).

The system will prompt you (see image below), where you can complete the project construction from zero to one, and all changes will be automatically saved, and two sample projects will be displayed: Snake and Todo List.

To test the normal process, I directly entered in the dialog box below: "Please generate a Todo List application with React."

The system will prompt you to specify a directory as the project directory. Then, it will continuously provide step-by-step instructions (see image below).

In the image above, when specific commands need to be executed, it will be accompanied by a "Run" button, asking you to click to confirm.
After clicking, Trae will automatically open the built-in terminal to execute the command (as shown in the image below).

Then, it will generate the required files one by one, requiring you to review them.

Once all files are generated, the command to start the local preview server appears.

After running it, Trae will launch a built-in webview page to display the rendering effects, updating in real time.

This is great; any changes can be seen immediately.
However, I must say that, perhaps due to the beta version, the Builder mode is currently not yet perfect. It is only suitable for generating project architecture and does not achieve a one-stop solution, falling short of the ideal state. I tried it multiple times, and if the requirements are complex, the generated code is difficult to run on the first try, requiring multiple modifications. People should have reasonable expectations for it, and this mode still needs further improvement.
VI. Other Features
Besides the two major modes, Trae's other features are also noteworthy.
(1) Code intelligence completion.
This is the basic function of the AI programming assistant, so I won’t elaborate on it. Just press Enter to create a new line, and Trae will read and understand the current code, then automatically complete the subsequent code.
If there are comments, it will generate the missing code based on the comments.
(2) Specify context.
Trae allows you to specify the scope of the conversation (context). The specific method is to enter the # number in the dialog box.

In the image above, after entering # in the dialog box, the system will automatically pop up a menu for you to select the context (context), with a total of four options.
- Code
- File
- Folder
- Workspace (Workspace)
The default context is the current workspace (Workspace, i.e., the currently open project), or it can be specified as a function or class (Code), a file (File), or a folder (Folder).
Once the context is specified, the AI's responses will be more targeted. For example, when generating code, it will consider the context of the scenario.
(3) One-click conversation transfer.
To conveniently transfer the content of the code editor to the AI, Trae provides a "One-click conversation transfer" button, saving the trouble of copying and pasting.

In the image above, selecting a segment of code will bring up a floating menu with "Add to conversation," allowing for one-click copying into the AI dialogue box for easy questioning.
Error messages in the terminal window also have this functionality (image below), making it very convenient for the AI to debug.

(4) Model customization feature.
I heard from a classmate at ByteDance that Trae will support model customization in the future, allowing users to connect the corresponding model API based on their preferences.
Seven, Summary
After trying the domestic version of Trae, I found it has several significant advantages.
(1) The product design is thorough, considering the most common scenarios, and the user experience is quite smooth.
(2) The interface is user-friendly, with good interaction design, making it easy for developers to get started.
(3) AI models (including the full-featured DeepSeek) are available for unlimited free use, with consistently stable and fast responses.
The drawback is that the Builder mode is not yet strong enough, and the generated programs have bugs. Often, the first generation doesn't run properly and requires continuous adjustments, which is time-consuming.
In summary, a complete AI IDE is still easier to use than an AI plugin. has a "sense of unity," achieving a higher degree of customization.
As AI capabilities continue to advance, the possibilities for AI IDE will expand further, potentially becoming the main direction for future IDE development.
The domestic version of Trae has just launched, and the development team hopes everyone will download and try it and provide valuable feedback.
(End)












