OpenAI Codex Tutorial: Code with AI in Minutes!

OpenAI Codex: A Cloud-Based Software Engineering Agent

OpenAI has recently launched Codex, a cloud-based software engineering agent. It's currently available to Pro, Enterprise, and Team users, with support for Plus and EDU users coming soon. This article explores the features and capabilities of Codex, particularly within a Team account.

Pricing and Access

• Plus Members: Codex is not currently supported.

• Pro Members: Access requires a $200 monthly subscription.

• Team Members: A monthly subscription costs $30 per user, while an annual subscription costs $25 per user (minimum of two users).

To experience Codex effectively, the Team option at $60 per month (for two users) offers a more cost-effective entry point compared to the Pro plan.

Codex Features and Capabilities

Once a Team account is activated, a Codex option appears in the left-hand menu. This provides access to the research preview version, which offers several features:

• Building Functionality: Codex can assist in building code.

• Codebase Understanding: It can answer questions related to the code library.

• Testing: Codex can run tests.

• Pull Request (PR) Drafting: It can draft PRs.

These functionalities can be executed in parallel, aligned with the official description of Codex's ability to "work on many tasks in parallel."

Codex can also run GitHub PR requests, lead code libraries, execute linting and testing procedures. These features are driven by a new code model known as Codex One, designed to adapt software development and optimize performance. The Codex One model is trained on actual code tasks, incorporating enhanced learning to generate code in a human style and produce PR code. Internal SWE Bench tests have demonstrated that the CodeX1 model boasts a higher accuracy compared to the O3 model.

Connecting to GitHub

Codex can be connected to a GitHub account and a specific repository. It prompts the creation of a basic environment for the repository, although this environment may not include all required dependencies, which can be added manually.

Task Types: Ask and Code

Codex offers two primary task types:

• Ask: Used for querying and understanding the code library (similar to the Chat function in Cursor).

• Code: Used for instructing Codex to write code (similar to the Agent function in Cursor).

Let's elaborate on each task type:

• Ask Example: Explaining the structure of a connected repository (e.g., LanChain) to a newcomer, highlighting key aspects and important content. This is particularly useful for onboarding new team members or understanding open-source projects.

• Code Example 1: Checking the code library for important bugs and fixing them.

• Code Example 2: Browsing the entire code library to identify issues, propose tasks for bug fixes, and suggest improvements to testing. This is analogous to a code review process.

Task Execution and Environment

Each task runs in an isolated environment pre-installed with the codebase. Codex can read files and execute commands within these environments. Task completion typically takes between 1 to 30 minutes, depending on complexity.

The environment can be configured with packages like Python, Node.js, Ruby, Rust, Go, and Java. Environment variables, secrets, and startup scripts can also be defined.

Crucially, Codex operates in a cloud-based, security-isolated container with no internet access during task execution. This ensures that the agent cannot access external websites, APIs, or services.

Analyzing Task Results

Successful tasks provide a summary and an execution log. The log details the environment setup, including pre-installed packages, and the execution of prompts. Codex searches for an Agents.md file, which (if present) should contain information about the project's structure, technologies, and conventions. If not found, it reads all files and folders to understand the project's functionality, structure, and key areas.

In "Code" tasks, the log shows the specific code modifications made, highlighting additions (green) and deletions (red). For example, Codex might adjust score keys to prevent shared states between strengths. It may also suggest setting up a starting script in the environment to install missing dependencies, as it may be unable to access the internet to install dependencies.

Pull Requests

Codex can automatically submit PRs with summaries of the changes and modifications to specific files. This offers a convenient way to review and integrate Codex's suggestions.

Conclusion

Codex can be a valuable assistant for software engineers, aiding in learning, coding, bug fixing, and testing. Its ability to review code and submit tasks in an isolated environment simplifies the development workflow. Consider Codex as an assistant, employee, or teacher who can make your development operations smoother.