OpenAI updated Codex: computer control, browser, memory, and plugins for developers

OpenAI has released a major update to the Codex desktop application for macOS and Windows. The new version introduces five key features: computer control (computer use), an embedded browser for in-app work, image generation, long-term memory between sessions, and plugin support. The goal of the update is to accelerate developer workflows and transform Codex from intelligent autocomplete into a full-fledged agent that independently executes tasks.

Computer use is the most significant of the new additions. Codex gained the ability to directly interact with the operating system: run terminal commands, manage the file system, open other applications, and perform complex sequences of actions without developer involvement. This is a fundamental shift: previously, Codex could only advise; now it can act. A developer needs only to formulate a task at a high level, for example "set up CI/CD for a new project," and the agent independently figures out the steps and executes them. This is especially important for DevOps tasks, which previously required a separate set of tools.

The embedded browser closes one of the chronic pain points in development. Programmers constantly have to switch between editor and browser: searching for up-to-date library documentation, checking ready-made solutions on Stack Overflow, reading the changelog of a new framework version. Codex takes this search upon itself: the agent independently finds the necessary information, analyzes it in the context of the current code, and immediately applies the result—without interrupting the workflow. This is particularly valuable when working with APIs, where backward compatibility often breaks between versions.

Image generation expands Codex's audience beyond backend development. Now you can create screen mockups, icons, or placeholder images directly in the working environment without switching to Figma or a separate image generator. For frontend teams, this eliminates the constant context gap between code and design. In early prototyping stages, where speed is more important than final quality, such integration saves time and removes unnecessary tool-switching overhead.

The memory feature changes the nature of interaction with the tool. Codex now remembers architectural decisions adopted in the project, code style conventions, peculiarities of the working environment, and the context of previous tasks. The longer a developer works with the application, the more precise the suggestions and actions of the agent become—this is a fundamental difference from stateless tools that start from scratch each time. Plugins complete the picture: developers will be able to connect corporate systems, their own tools, and external services to Codex, customizing the workflow to the needs of a specific team. Each connection expands the agent's capabilities without needing to change the main application.

This update clearly marks OpenAI's strategic direction: from language model to agentic working environment. Codex now competes not only with GitHub Copilot, but also with a new generation of tools—Cursor, Windsurf, Devin. The company is consistently shifting the product from the "assistant" category into the "autonomous executor" category. Developers should monitor this transformation: tools that seem cutting-edge today risk becoming the standard in a year. Those who master agentic workflows now will gain a real advantage.