Video tutorial · The Metaverse Guy
I Tested OpenAI Codex on a Real Flutter App... Here's What Happened
I put the new OpenAI Codex extension in VS Code to the test on a Flutter fitness app, integrating an AI chatbot with the GLM 4.6 model via OpenRouter.
- Published
- 17 October 2025
- Length
- 6 min
- Views
- 9,369 views
OpenAI recently released a new version of Codex that works directly inside IDEs like VS Code. I tested it on a larger Flutter codebase—a fitness app originally built with Augment Code—by asking it to integrate AI chat functionality using the GLM 4.6 model through OpenRouter. The test revealed Codex's strengths in context engineering and automated coding, though speed could be improved compared to alternatives like Claude Code.
What I Tested
OpenAI just released a new version of Codex that you can access directly inside your favourite IDE, much like Claude Code, Cline, or Roo Code. I wanted to see how it performs on a real-world project, so I challenged it to understand an existing Flutter application and add AI functionalities.
The app was a fitness application originally built using Augment Code. My goal was to integrate an AI chatbot feature using the GLM 4.6 model through OpenRouter, all while working inside VS Code with the new Codex extension.
Setting Up Codex in VS Code
Getting started with Codex in VS Code is straightforward. After installing the new Codex extension, I configured it to work with my Flutter project. The setup process gives you access to agent mode, which allows Codex to read, write, and edit files across the codebase.
I didn't need any complex configuration—just pointed it at the project and started prompting. The auto-context feature helped it quickly grasp the existing code structure.
Integrating AI Chat with GLM 4.6
The main task was to add an AI chatbot to the fitness app. I used OpenRouter to connect the GLM 4.6 model, and Codex handled the integration from prompt to code. It generated the necessary Dart code, wired up the API calls, and placed the new feature within the existing UI.
Along the way, I had to do some prompt engineering to guide it, but Codex managed to understand the app's architecture and make the changes without breaking existing functionality.
How Codex Performed
Codex showed real strength in automated coding and context engineering. It could navigate a larger codebase, understand dependencies, and generate code that fit the project's style. Bug fixing was also handled reasonably well when issues came up.
However, speed was a noticeable drawback. Compared to some other AI agents, Codex took longer to process requests and apply changes. For a production-ready app, that extra time could add up, but the quality of the output was solid.
Codex vs Other AI Coding Tools
While this test focused on Codex, I couldn't help but compare it to tools like Claude Code, Cline, and Cursor. Codex's context engineering felt more thorough, which is great for complex tasks, but the slower iteration speed might frustrate developers who value rapid feedback.
For automated coding and understanding an existing codebase, Codex holds its own. It's a promising addition to the AI coding toolkit, especially if OpenAI continues to improve its performance.
Tools used in this video
- OpenAI Codex
- VS Code
- Flutter
- Dart
- OpenRouter
- GLM 4.6
- Augment Code
- Claude Code
- Cline
- Cursor
Questions
- How do I set up OpenAI Codex in VS Code?
- Install the new Codex extension from the marketplace, then open your project. The extension provides agent mode with full access to your codebase, and the auto-context feature helps it understand your project structure immediately.
- What is the GLM 4.6 model and how does it integrate with Codex?
- GLM 4.6 is a language model accessible via OpenRouter. In this test, I used Codex to write the integration code so the Flutter app could send and receive messages through the OpenRouter API, effectively adding an AI chatbot to the fitness app.
- How does Codex compare to Claude Code or Cursor?
- Codex excels at context engineering and automated coding, making it strong for understanding larger codebases. However, it was slower in this test compared to Claude Code and Cursor, so there's a trade-off between depth of understanding and speed.
- Can Codex handle large Flutter codebases?
- Yes, I tested it on a full fitness app that was originally built with Augment Code. Codex was able to navigate the existing code, add new features, and fix bugs without losing track of the project's structure.
Topics
Video description and links
This article was written from the video's description with the help of AI. Watch the video for the full walkthrough.



