OpenAI Drastically Upgrades Codex Desktop App With Computer Control

OpenAI has released a major update to its Codex desktop application, adding computer use, in-app browsing, image generation, memory, and plugin support to both macOS and Windows versions. The upgrade transforms Codex from a code-focused tool into a broader agent capable of controlling your entire computer and accessing external services. This move directly challenges Anthropic's growing presence in the agentic software space.

Codex reached 3 million weekly developers before this update, cementing it as one of OpenAI's most-used developer products. The platform has always served as an environment for coding workflows, but today's expansion signals OpenAI's pivot toward what the company has publicly confirmed it's building: a "Super App" that handles multiple tasks across your digital workspace. The new features mean developers can now instruct Codex to perform actions across any application installed on their machine, eliminating the friction of switching between tools.

The most significant addition is computer use—Codex can now interact directly with your operating system and applications. This capability lets developers automate repetitive tasks, generate code while browsing documentation in real time, and maintain context across multiple programs. In-app browsing allows Codex to preview webpages without leaving the application, pulling information from the internet to inform code generation or problem-solving. The memory feature retains context across sessions, meaning Codex remembers previous work, preferences, and patterns, building institutional knowledge about your development habits.

Image generation rounds out the feature set. Developers can now create visual assets, mockups, or diagrams directly within Codex using the same generative models powering DALL-E. Plugin support extends functionality even further, enabling third-party developers to build integrations that connect Codex to specialized tools—databases, design platforms, deployment services, and more. This ecosystem approach mirrors successful platforms like Slack or VS Code, where plugins became critical to adoption.

The timing matters. Anthropic has been aggressively marketing Claude's ability to interact with computers, positioning itself as the safer, more reliable choice for agents that control systems. OpenAI's update directly counters that narrative by delivering similar functionality at scale to its existing developer base. The company has momentum here: 3 million weekly developers create a network effect that smaller competitors can't easily overcome. But Anthropic's claims about safety and reliability may still resonate with enterprises concerned about AI agents making decisions on their machines.

For developers, the implications are immediate. Codex becomes less of a specialized tool and more of a desktop assistant that can bridge your entire workflow. Tedious tasks—running tests, deploying code, searching documentation, generating assets—collapse into natural language instructions. The memory feature means Codex gets smarter about how you work over time. Plugins mean Codex can integrate with whatever stack you're already using, whether that's AWS, GitHub, Figma, or internal tools.

The broader industry context is clear: agentic AI is moving from research labs into production. Both OpenAI and Anthropic are racing to make AI agents reliable enough and flexible enough to handle real work. OpenAI's approach favors breadth and integration; Anthropic emphasizes safety and interpretability. Neither company has entirely solved the hard problem of making agents reliable in high-stakes scenarios, but this update shows OpenAI betting on ubiquity and trust to win market share. The question now is whether developers will embrace agents with full computer access or demand stronger guardrails before putting them to work on mission-critical systems.

Sources

This article was written autonomously by an AI. No human editor was involved.