OpenAI Expands Codex Desktop Capabilities for macOS

OpenAI has evolved its Codex desktop application from a sandboxed coding environment into a sophisticated agent capable of controlling macOS applications and executing background tasks [2]. The platform now utilizes the GPT-5.5 model to perform complex workflows, allowing users to automate software interactions and manage local files directly from their Mac [1, 2].

Key takeaways

• The updated Codex app features a "Computer Use" capability that allows the agent to operate the mouse and keyboard within any Mac application [1, 2].
• A new feature called Chronicle captures periodic screenshots to create text-based memories, enabling the AI to maintain context across different sessions [2].
• OpenAI introduced a mobile integration that allows users to monitor and control active desktop Codex sessions from their smartphones [2].
• The agentic software is powered by the GPT-5.5 model, which achieved an 82.7% accuracy rate on the Terminal-Bench 2.0 coding benchmark [2].
• Security documentation warns that Chronicle stores unencrypted Markdown memory files locally, which are accessible to other applications on the machine [2].

From Sandboxed Code to Desktop Agent

The transformation of Codex occurred through a series of updates this spring, shifting the tool from an isolated cloud environment to a desktop-integrated agent [2]. By adding computer-use capabilities, OpenAI enabled the software to navigate third-party applications, fetch data from browsers, and install plugins from a catalog of over 90 options, including integrations for Jira and Microsoft Suite [2]. These "Skills" can be triggered on automated schedules, functioning similarly to a natural-language personal cron job [2].

To address the challenge of AI losing context between sessions, OpenAI released the Chronicle feature [2]. This system uses optical character recognition to summarize screen activity into Markdown files stored on the user's device [2]. While this allows the agent to recall past work, it has drawn scrutiny from security experts due to risks like prompt injection, where the AI might inadvertently follow instructions hidden on a webpage it has captured [2]. OpenAI advises users to pause the feature when viewing sensitive information [2].

Mobile Control and Future Outlook

In May, OpenAI expanded the reach of its desktop agent by allowing the ChatGPT mobile app to serve as a remote control [2]. Users can view terminal outputs, file changes, and live screenshots of their desktop sessions from their phones, though the mobile app itself does not execute the code [2].

For many users, the ability to build custom tools has become a primary use case for the desktop app [1]. By utilizing the "Computer Use" feature, individuals have successfully created personalized applications—such as activity trackers or customized social media interfaces—without needing extensive experience in Xcode [1]. As OpenAI continues to integrate these capabilities, the company has indicated that some of this functionality may eventually be absorbed into the main ChatGPT application, though the specific timeline for such a transition remains unclear [1].