What Is Codex? OpenAI's One Agent for Everywhere You Code

TL;DR

Codex is now one software engineering agent connected by your ChatGPT account, usable in an IDE extension, CLI, GitHub code review, web, and the iOS app. It runs either locally on your machine for pair programming or remotely in a secure cloud sandbox that returns a pull request. It is included with Plus, Pro, Business, Enterprise, and Edu subscriptions.

Transcript

Welcome everyone uh to another episode of Build Hours. I'm Pranadesh Mande. Hi, I'm Dominic Kundal. And we're your hosts for today as we dive deep into Codeex, our software engineering agent. Before we get into the content, just a quick plug for build hours. This is our live virtual event series for builders and developers to help you get the most ... Read More

Key Insights

• Codex is now one agent for everywhere you code, connected by your ChatGPT account, so signing in with ChatGPT in any interface lets you use the same functionality across the IDE extension, CLI, GitHub, web, and the ChatGPT iOS app.
• The Codex IDE extension works in VS Code, cursor, and any VS Code-compatible fork, bringing the same functionality as the CLI directly alongside your code so you can pair program more easily.
• Codex runs in two places: locally on your machine for pair programming where it executes code and modifies your environment, or remotely in a secure cloud sandbox that works asynchronously, downloads your code, and returns a pull request.
• The new mental model for Codex has two dimensions: where you as a developer use it (IDE, terminal, GitHub, web, mobile) and where the actual work runs (locally or in the cloud sandbox).
• Code review is a new feature that gives automatic code reviews for new pull requests once enabled in GitHub.
• Codex is included in ChatGPT subscriptions including Plus, Pro, Team (now Business), Enterprise, and Edu, requiring only a ChatGPT sign-in rather than separate API keys.
• Structuring a repository clearly helps Codex handle tasks, since a monorepo with clearly named projects lets the agent traverse packages and find answers the same way a human engineer would.
• A novel new feature lets you hand off tasks to the cloud while working and pull them back down, making Codex feel like an agent available everywhere you code.

Questions & Answers

Q: What is Codex and how is it structured now?

Codex is OpenAI's software engineering agent, now brought together to feel like one agent for everywhere you code, connected by your ChatGPT account. It is structured along two dimensions: where you as a developer use it and where it actually runs. You can use it through an IDE extension, the CLI, GitHub for code reviews, the web, and the ChatGPT iOS app, while the work itself runs either locally on your machine or remotely in a secure cloud sandbox.

Q: Where can I use the Codex IDE extension?

The Codex IDE extension works in VS Code, in cursor, and in any fork that is compatible with VS Code, such as wind surf. It brings the same functionality of the CLI into your IDE so you can work with Codex alongside your code more easily. You can find it in the extension marketplace on VS Code or cursor, and it appears on the side panel where you can start asking questions about your repository.

Q: How does Codex run tasks locally versus in the cloud?

Codex runs in two places. Locally, it works on your machine, which is more like pair programming where it executes code and modifies the environment on your machine. Remotely, it runs asynchronously in a secure cloud sandbox on its own machine, downloads your code, and does a bunch of work for you to give you a pull request that you can either merge or continue working from.

Q: Is Codex included in ChatGPT subscriptions?

Yes. If you are on Plus, Pro, or Team, which is now Business, Enterprise, or Edu, Codex is available to you. It is part of the subscription and is connected by your ChatGPT account. You just sign in with ChatGPT in any of the interfaces, such as the IDE extension, CLI, web, or iOS app, and you can start using it without needing separate API keys.

Q: What new features did Codex add last week?

Last week's release rolled out a slew of new features. These included a new IDE extension that works in VS Code and compatible forks, continual improvements to the Codex CLI with releases every few days, and code review, which provides automatic code reviews for new pull requests when enabled in GitHub. It also added the ability to hand off tasks to the cloud while working and pull them down later.

Q: Why was there confusion around the Codex naming?

Confusion was understandable because many things were named Codex. There was the Codex CLI and Codex in ChatGPT, which didn't work together, plus Codex one and Codex one the model, and also the Codex model from 2020. Simplifying that naming and the mental model has been a big focus, so that now everything is connected via the ChatGPT account and Codex feels like one product with one name, similar to how GitHub works across CLI, web, and mobile.

Q: How does structuring a repository help Codex?

Structuring a repository can be extremely helpful for how Codex deals with tasks. In the demo, having a monorepo with clearly named projects is incredibly helpful because it allows Codex to traverse the different packages to find answers. When asked which real-time API example is most comprehensive, Codex has to search files, read additional files for context, and compare the level of features implemented across packages, the same way an engineer would.

Q: What model does Codex use in the IDE chat demo?

In the IDE extension demo, working locally in chat mode, Codex uses GPT-5 on medium reasoning, which you can change in the interface. Chat mode means it answers questions like what the repo is about using the local context of the repository, traversing the codebase the same way an engineer would. The demo was run on the open-source OpenAI Agents SDK for TypeScript repository. GPT-5 was launched only a month before this event as OpenAI's best scoring model.

Summary & Key Takeaways

• Codex was first released in April as a lightweight open-source CLI coding agent usable with ChatGPT or API keys, followed by Codex in ChatGPT, an asynchronous cloud agent that connects to GitHub, runs code remotely, and returns a pull request. These two experiences were powerful but siloed and did not work well together.

• Last week's release brought the experiences together so Codex feels like one agent for everywhere you code. It added a new IDE extension for VS Code and compatible forks, a continually improved CLI, automatic GitHub code reviews for new PRs, and the ability to hand off tasks to the cloud and pull them back down.

• The demos use Codex to maintain the open-source OpenAI Agents SDK for TypeScript. Working locally in chat mode, Codex uses GPT-5 on medium reasoning to answer questions like what the repo is about and which real-time API example is most comprehensive, traversing the monorepo and comparing packages the way an engineer would.