OpenAI is releasing a new desktop application . This allows developers to control multiple AI agents simultaneously during software development.
This Codex app is currently available for macOS. It is intended to provide a central working environment for tasks performed in parallel by different agents.
According to OpenAI, the tool aligns with a changing way of working in which developers no longer use a single assistant but instead deploy multiple agents to perform long-term tasks independently.
Since the introduction of Codex, the role of AI in development processes has shifted, according to the company. Whereas agents were previously used primarily for specific code adjustments, they are now used for broader processes that span the entire development lifecycle. Developers delegate tasks, run processes in parallel, and entrust agents with projects that can take hours or even days to complete. According to OpenAI, existing IDEs and terminal tools are not designed for this, creating a need for a separate environment for such collaboration.
Agentic software development is a trend
With this app, OpenAI also seems to be joining a broader trend of so-called agentic software development, in which AI agents perform programming tasks largely independently. Tools such as Claude Code and Cowork are often cited as pioneers in this approach.
The Codex app is designed to fulfill this role by allowing agents to work in separate project threads. Users can switch between tasks without losing context and immediately review changes made by agents. The application supports worktrees, allowing multiple agents to work on the same repository without causing conflicts. Each agent works in an isolated copy of the code, allowing different solution directions to be explored in parallel without affecting the local git status.
An important part of the new approach is so-called skills. These enable Codex to do more than just generate code. Skills combine instructions, scripts, and access to external tools, allowing the agent to perform tasks such as retrieving design files, managing project information, or generating documentation. Within OpenAI, such extensions have long been used to automate recurring tasks, ranging from analyzing error messages to preparing reports.
Flexible interface
According to OpenAI CEO Sam Altman, the underlying model, GPT-5.2-Codex, is currently the company’s most powerful model for programming tasks, but until now, it has been less accessible to use. The new app aims to lower that barrier by linking the model’s capabilities to a more flexible interface.
That claim deserves some nuance. TechCrunch reports that while GPT-5.2 ranks at the top of TerminalBench, models such as Gemini 3 and Claude Opus achieve comparable scores within the margin of error. No clear advantage is visible in SWE-bench either. At the same time, these types of agentic applications are difficult to measure unambiguously, and differences in user experience play a major role.
In addition, the app introduces support for Automations. This allows Codex to perform tasks in the background according to a fixed schedule. Results appear in a queue where developers can review them later. OpenAI sees this as a way to reduce repetitive work without constant supervision.
The application includes standard system-level sandboxing. Agents are only allowed to work within the assigned folder or branch and explicitly request permission for actions with elevated privileges, such as network access. Teams can customize these rules to suit their own working methods.
The Codex app is now available for macOS users with a paid ChatGPT subscription and, temporarily, for users of the free version. According to OpenAI, a Windows version is in the works.