Write Code in IDEA with Codex—Zero Manual Typing Required

JetBrains IDEs ship with an AI assistant that often feels slow and struggles with complex logic, prompting many developers to seek better solutions. On January 22, OpenAI released a Codex extension that fully supports the JetBrains family (IntelliJ IDEA, PyCharm, WebStorm, Rider, etc.). Starting with version 2025.3, the built‑in Assistant can bind a ChatGPT account, offering a smoother experience.

Configuration is straightforward: open the Assistant sidebar, click “Sign in Codex with ChatGPT”, install the plugin if prompted, choose “Log in via ChatGPT account”, and complete the authorization in the browser. The process finishes in a few clicks.

The extension runs the latest GPT‑5.4 model and incorporates a GPT‑5.3‑Codex capability tailored for code. On the SWE‑Bench Pro benchmark it achieves a 57.7 % score, comparable to the original Codex’s 56.8 % but with lower latency, making code generation feel more responsive.

Practical demos show Codex handling real development tasks. For example, a single command can download a GitHub project (e.g., https://github.com/rymcu/mortise) and set it up even when Git is not installed, with Codex automatically installing the missing tool. When presented with an unfamiliar codebase, Codex can explain each module, list the tech stack, and outline the project structure, while also detecting and installing required tools such as JDK 21, PostgreSQL, or Redis, and generating necessary scripts.

Beyond project bootstrapping, Codex excels at global refactoring: it can scan an entire codebase, prioritize optimization points, and modify dependent files (e.g., updating all call sites after an interface signature change). It also auto‑writes and runs tests, generating both normal and edge‑case scenarios, and iteratively fixes failures until all tests pass. Bug diagnosis is performed directly in the project context, providing solutions that are specific to the code rather than generic web search results.

The author, with modest front‑end skills, built a small application that calls an AI service to generate stylized couple avatars, handling login, AI image generation, cloud storage, membership, and payment logic—all without manually writing code. UI components were generated with Bolt and fine‑tuned by Codex.

Finally, the author notes that while Claude offers similar capabilities, it is highly sensitive to network conditions and can be blocked easily, whereas ChatGPT remains more stable for daily use.