Which is a Better Coding Agent?

Open-source CLI tools have created a lot of excitement in the developer and AI communities. With the release of Claude’s Codex CLI, OpenAI’s Codex CLI, and, more recently, Google’s Gemini CLI, there is a sudden increase in interest in CLIs in general. By incorporating AI support into routine tasks, these tools are changing how developers use their terminals.

Developers can improve code comprehension, speed up debugging, and automate difficult coding tasks by utilising these CLIs. Tools like Codex CLI help users write, analyse, and improve code more efficiently by bringing strong AI-driven features like code generation, refactoring, and inline documentation directly into the command line. The market has become even more exciting with the launch of Gemini CLI, a lightweight, open-source substitute from Google that is gaining popularity quickly.

Let’s now take a closer look at how Gemini CLI and Codex CLI fare when put to the test.

What are Gemini CLI and Codex CLI?

Gemini CLI: An open-source AI agent by Google that integrates Gemini models (like Gemini 2.5 Pro) into the terminal for developers.

Codex CLI: An open-source command-line tool forked from OpenAI’s original Codex CLI, now supporting multiple AI providers, including Gemini 2.5 Pro. It enables developers to interact with large language models in the terminal for tasks like code generation, refactoring, and script explanation.

If you’re looking for detailed walkthroughs, check out our guides on each CLI tool:

Installation

Before proceeding with the installation, ensure that the system requirements are satisfied.

System Requirements

NodeJS is a common prerequisite for both the CLIs. You can install Node.js by following the instructions on its official page: https://nodejs.org/

Installation Steps for Codex CLI

1. Check Node.js version:

node --version

2. Ensure it is NodeJS version 22 or higher.
3. Install Codex CLI globally:

npm install -g @openai/codex

4. Or using pnpm:

pnpm add -g @openai/codex

5. Set your OpenAI API key (if within your project directory’s .env file contains the OPEN AI API Key, then we don’t need to do this task):

export OPENAI_API_KEY="your-openai-api-key"

6. Add this line to your shell config file (e.g., ~/.bashrc, ~/.zshrc) for persistence.
7. Run Codex CLI:

codex

8. Or pass a prompt directly:

codex "explain this codebase to me"

Installation Steps for Gemini CLI

1. Check Node.js version:

node --version

Ensure it is 18 or higher.

2. Install Gemini CLI globally (recommended):

npm install -g @google/gemini-cli

3. Alternatively, run directly without installation:

npx https://github.com/google-gemini/gemini-cli

4. Authenticate:
   1. Default: Sign in with your Google account when prompted (browser window will open).
   2. API Key (for advanced use):
      1. Get your API key from Google AI Studio.
      2. Set it in your environment using the following entry:
         export GEMINI_API_KEY=”your-api-key”
5. Run Gemini CLI:

gemini

6. Or pass a prompt directly:

gemini "Hello, can you help me with coding?"

Things to Keep in Mind

• For both tools, Node.js and npm (Node Package Manager) must be installed and available in your system PATH.
• For Codex CLI, Node.js 22+ is required, while Gemini CLI works with Node.js 18+.
• Both require authentication via API keys or account login for full functionality.
• On Windows, Codex CLI is best run via WSL2 for full compatibility.
• Remember, when you want to use your CLI for a certain project, make sure you have the current working directory set to the folder path where your project exists. 

Running our CLI Instances

Let’s proceed to compare the performance between the 2 models with 3 tasks

• Building a working model on a local host
• Reading a complex codebase and explaining the code
• Error debugging

Let’s first try it out with Codex CLI:

Codex CLI

If all the requirements have been satisfied, then we can start with the tests.

Task 1 – Explaining our Codebase

We can see that Codex can provide a detailed summary of our code base, which was, to be honest, surprisingly good. Additionally, Codex CLI temporarily asks us to grant access to specific code segments, which we can grant or reject based on our judgment. It provides a respectable report after taking a reasonable amount of time to reflect.

The report’s high-level purpose, repository layout, key components, front-end websites, and assets are all included, along with a summary. Even though I had given the main files rather ambiguous names, I was still shocked to find that it could decode all of the previous iterations of the code files in chronological order.

Task 2 – Error Debugging

Since we allow the Codex CLI access to our entire database, whether it be private or protected, it will ask for our authorization before reading any code snippets. If we permit it, it will access our snippet appropriately. 

Overall, it was able to recognise every minute mistake and occasionally offered fixes for some careless errors in the code, removing unnecessary portions that had an impact on some of the latency slightly.

Task 3 – Additions to Codebase

This involved understanding how each module was defined across various directories and bridging the knowledge between each dependency to ensure that a new directory was created to build the new.py files for the additions we had requested when we requested access.

Let’s now attempt to examine how the Gemini CLI functions.

Gemini CLI

I liked the Dracula theme.

Unlike Codex CLI, where we only need to export our OPENAI_API_KEY, we have three different ways to log in here: using our Gmail account, copying the GEMINI_API_KEY from AIStudio, or using Vertex AI credentials.

Task 1 – Explain our Codebase

The main difference I observed was that, unlike Codex CLI, which asks for access to specific directories, Gemini skips this step and instead scans our entire database. Nevertheless, in my experience, it does offer a decent overview, but it is by no means as good as Codex in terms of output.

Task 2 – Error Debugging

In my opinion, Gemini’s response was significantly better explained than Codex’s, as it was able to identify the kinds of security risks, explain them, and provide file names as references for further investigation and suggested changes. Additionally, offering a thorough strategy to address each of these mistakes would enable the code to be modified without impairing its functionality if the necessary instructions were followed.

Task 3 – Additions to Codebase

Gemini CLI wasn’t working for this workload, so no output was produced.

Gemini CLI vs Codex CLI: Final Verdict

Also Read: 10 Ways Students Can Use Cursor AI for Free

Conclusion

When comparing the features of these two CLIs, I found that Codex offers more functionality than Gemini CLI. At the moment, both the CLIs are open-sourced and will eventually have better functionality in future releases. 

Since a GPT model analyses our code, Codex also offers greater privacy by requesting access to our codebases and modules in a way that minimizes security risk. However, this doesn’t appear to be a case of the Gemini CLI requesting access regularly; instead, it asks when it wants to run the application code. Additionally, we should remember that the context on the right side of the terminal is also tracked by Gemini CLI. 

When combined with more intelligent vibe coding IDEs like Cursor, Windsurf, and others, such CLIs would undoubtedly open up new possibilities for vibe coding workflow and application operation. However, security and privacy vulnerabilities may still exist, so I would advise being careful about what and where in our codebase we grant these CLIs access.  

Frequently Asked Questions