I. Introduction

This week, I plan to write an additional article.

This is because AI programming IDEs have suddenly become popular, with significant events happening both domestically and internationally.

First, let's talk about what's happening abroad. OpenAI is planning to acquire Windsurfwith $3 billion..

This news is astonishing. Windsurf (formerly known as Codeium) has a very short history, having been released only two years ago, and its market share remains relatively low. Yet, it's valued at so much!

OpenAI's latest funding round (March of this year) was only$4 billion, yet now it's suddenly willing to spend $300 million to acquire it. What exactly is it looking for in Windsurf?

OpenAI doesn't have its own programming assistant, so the only explanation is that it wants to acquire an IDE to enter the AI programming market, which is very important to it.

II. MarsCode Renamed to Trae

Looking at China, ByteDance is also making big moves.

Its programming assistant was originally the MarsCode plugin, and later it added an independent AI IDE product Trae.

This week, MarsCode announcedthat it has been renamed to Trae plugin, and it will no longer be developed as a standalone brand.

Going forward, ByteDance's AI programming assistant will only have the Trae brand. comes in two product forms.

for users accustomed to traditional IDEs, can install the Trae plugin; users seeking a better AI experience can install the standalone Trae IDE.

while this announcement was made, the new version of Trae was also released, adding significant new features (more details to follow).

it can be seen that ByteDance has made a firm decision to integrate the products, preparing to focus on AI programming tools.

III. AI IDE + MCP

why do giants both domestically and internationally, at the same time, all turn their attention to AI IDE?

I speculate the answer is the emergence of MCP.

with MCP, AI IDE can extend external capabilities, making it all-encompassing, which makes it a must-compete space for giants.

Below, I will explain what MCP is and how to call it in Trae. After reading this, you will understand why MCP is so important.

Four, an introduction to Trae

I chose Trae for demonstration mainly because it is a domestic software with a Chinese interface and documentation, and it is completely free (foreign products all require payment).

As mentioned before, Trae comes in two forms: plugins and IDEs. Its IDE is divided into a domestic version and an overseas version. These products have almost the same functions, just with different built-in AI models.

Domestic version: Built-in deepseek R1, V3, v3-0324, and doubao 1.5 pro models

Overseas version: Built-in Claude 3.5, 3.7, Gemini 2.5 pro, GPT-4o, and GPT-4.1 models

I suggest using the domestic version because the built-in models of the overseas version often require queuing, which is a waste of time, and it may also result in poor communication.

However, both versions support custom models , you can provide a key to access the model you specify. So, the difference between the versions is not very important.

By the way, the word "Trae" means "The Real AI Engineer" (a real AI engineer). I used to always think that "Trae" meant "True AI".

Section 5: Trae's New Version

Trae's MCP calling feature was introduced starting from version v0.5.0.

For those who haven't installed it, you can download the new version from theofficial website. For those who have already installed it, please check the version.

The interface has been simplified this time; the chat box and Builder (Project Generator) have been combined into a single dialog box (see image below). All interactions with AI are input here.

In the image above, there are two new buttons in the lower left corner: "@Agent" and "#Context." These are the core features added in this update.

As for the basic usage of Trae, it won't be mentioned here. You can refer toprevious articles.

Section 6: Calling the Agent

The entry point for MCP calls is the "@Agent" button in the lower left corner of the image above.

If you want to expand the functionality of AI, you need to use this button. Because the essence of an AI model is just a language model, its own functionality is limited and must be extended through external applications (agents).

Click "@Agent" (or type@），a dialog box will pop up displaying the available intelligent agents (as shown in the image below).

It can be seen that Trae comes with two built-in agents: "@Build" and "@Builder with MCP".

Among them, "@Build" is used to let AI generate a new project that can be run.

@Build Tetris Web Game

Entering the above command will generate an HTML file. Opening it will present a Tetris game.

Another built-in agent "@Build with MCP" is used to connect the MCP server.

Section 7: What is MCP?

First, let me explain what MCP is; it's quite easy to understand.

We know that AI models extend their functionality by connecting to external applications. Each external application has a different interface, so if you want to integrate 10 applications, you would need to write 10 different integration codes, which is very cumbersome. Moreover, if you switch to a different model, you might have to rewrite all the integration codes.

Given this, Anthropic proposed the MCP protocol in November 2024. External applications only need to support this protocol and provide an MCP interface (also known as an MCP server), so AI models can integrate using a unified format without needing to understand the integration details of the external applications.

Therefore, MCP can be understood as an adapter layer between AI and external applications.。For AI, as long as the MCP server of a certain app is installed, it can access the app without writing any code (except for a few configuration items).

Since MCP solves the access pain point for AI apps, it has become extremely popular in just half a year since its inception, even with OpenAI, a competitor of Anthropic, publicly supporting it. There are already over 10,000 open-source MCP server projects online.

Section 8: Calling MCP

Now let's see how Trae calls MCP.

Click the gear icon in the upper right corner of the AI tab to bring up a menu, and select the MCP option.

It will bring up an MCP tab (see image below), click the "+ Add MCP Servers" at the bottom.

Trae has a built-in MCP marketplace, offering some commonly used MCP servers. If you can't find the one you need, you can click "Manual Configuration" to add your own MCP.

To facilitate the demonstration, I chose the first server Puppeteer, allowing the AI to invoke the headless browser.

Clicking on the name of Puppeteer will take you to the homepage of this open-source project, where you can review the internal commands (i.e., capabilities) it provides.

As shown in the image above, this MCP server offers internal commands such as puppeteernavigator (to open a specified URL), puppeteerscreenshot (to take a screenshot), and puppeteer_select (to select page elements), which can be invoked by the AI model.

Users do not need to memorize these commands; they only need to understand the capabilities it offers.

Next, click the plus sign next to it to add this MCP server.

This MCP is labeled with "Easy Configuration," indicating that no setup is required and it can be run directly.

All self-added MCPs are placed in the built-in agent "@Build with MCP" by default, so they can be used through this agent.

In the AI dialog, select the agent "@Build with MCP" and then enter the command "Open https://www.baidu.com" to test the newly installed Puppeteer server.

Normally, Trae will let you choose a project folder, and then it will open a browser window displaying Baidu's homepage.

This is the function of MCP. AI originally did not have the ability to control the browser, but now it can be controlled through MCP.

Next, you can give more complex commands, such as generating screenshots, which can also be completed smoothly.

This is the basic process of calling MCP. You can also save the added MCP server as an agent (see the image below).

Then, by calling the agent with the name you set (as shown in the figure below), connect to the designated MCP server.

Section 9: Contextual Features

In addition to MCP calls, this update to Trae has also enhanced contextual features, which are briefly mentioned here.

Context refers to additional information provided to help the AI model think and complete tasks.

By #, you can access the contextual menu.

As shown in the figure above, the available contexts include additional code (code), files (file), directories (folder), and workspaces (workspace).

This update adds two more options, "Doc" representing additional documentation.

Clicking "Add Document Set" allows you to add a document directory as the context for the AI model.

Another option "Web" indicates using online information as context. This provides real-time connectivity for AI.

Regarding the real-time weather issue in the figure above, AI can only answer it with connectivity.

Ten, Summary

With MCP calls and connectivity, AI IDE has immense potential, not just as a programming tool but as a versatile AI console.

Major companies must have seen this potential, hence their willingness to invest significant resources into this product.

I believe that calling MCP servers within AI IDE will become a hot topic in the software industry in the near future, worthy of everyone's attention.

(End)