OpenAI turns Codex Desktop into a PC agent with a browser, memory and plugins

OpenAI is effectively rebuilding Codex Desktop from a code-writing tool into a control layer on top of the work computer. After the major update, the app can do more than help with development: it can see the interface, click, type, work with a browser inside the app and carry on long tasks without constant user involvement.

OpenAI presented the update on April 16, 2026. According to the company, more than 3 million developers already use Codex every week, and the new release is clearly designed to keep them inside a single workspace. Instead of a mix of IDE, terminal, browser, task tracker and separate automations, OpenAI offers one hub where you can delegate not only code to the agent, but also the surrounding actions. ZDNet describes this as an attempt to turn Codex from a coding tool into a full-fledged work hub. The desktop app is already available on Mac and Windows, but the most ambitious features are rolling out gradually.

The main addition is a background computer-control mode. Codex can now work with apps on the device using its own cursor: it sees the screen, presses buttons and types text. OpenAI separately stresses that several agents can run in parallel on a Mac without interfering with what the user is doing at that moment. The practical scenario is straightforward: a frontend developer tweaks an interface while the agent opens the app in parallel, checks the changes, repeats test actions and returns the result. This matters especially for programs that have no proper API, where automation previously stalled on manual work.

The second notable update is the built-in browser. Inside Codex you can now open pages and leave comments directly on interface elements to explain more precisely to the agent what needs fixing. For now OpenAI positions this mainly for frontend work, apps and games, especially when you need to move quickly through the cycle of "look at the screen, give an instruction, get a new version." Codex has also been taught to work with image generation via gpt-image-1.5, so a single flow can cover assembling a mockup, writing code and generating the visuals you need.

The third layer of changes concerns the developer workflow. Codex Desktop can now process comments from GitHub review, keep several terminal tabs, connect to remote devbox environments over SSH in alpha mode, and show previews of PDFs, spreadsheets, slides and documents in the side panel. There is also a summary panel for tracking the agent's plan, the sources it used and artifacts.

OpenAI also added more than 90 new plugins that combine skills, app integrations and MCP servers. The point is not a nicer catalog: it lets the agent pull context and take actions in JIRA, CI/CD, documents, cloud services and other work tools without constantly switching between windows.

Automations and memory are an equally important part of the release. Automations can now be attached to existing threads, so Codex picks up accumulated context instead of starting the task from scratch on each run. The agent is also allowed to schedule work for the future: it can "wake up" later and continue a long task days or even weeks on.

In preview mode there is now memory that stores user preferences, corrections and information that previously took time to gather. On top of this, Codex has started suggesting where to pick up your day: for example, pointing to unresolved comments, pulling in related context from connected tools and producing a list of next actions. OpenAI notes that computer control is available only on macOS at launch, and some personalized features are still catching up for enterprise users and certain regions.

What does this mean in practice?

Codex Desktop does not yet look like a replacement for traditional software in the literal sense. It does not eliminate IDEs, browsers, task trackers and specialized apps; rather, it is starting to act as a coordination layer above them. But the direction is clear: OpenAI is moving the product from an "assistant that writes code" to an "agent that runs the workflow." If computer control proves reliable, and memory and plugins do not break on real tasks, Codex could become not just another AI tool in the stack, but the interface through which the whole stack is used.