OpenAI Codex Update

On April 17, OpenAI announced an update to Codex, introducing several new features tailored for Mac users, including cursor-level background interaction, a built-in in-app browser, integration of gpt-image-1.5, and over 90 new plugins. Additionally, the automation capabilities of Codex have been upgraded, enhancing its memory function to support the reuse of conversation threads and context. Codex can now autonomously schedule and wake up to execute long-term tasks, proactively providing users with suggestions for subsequent actions based on context.

The highlight of this Codex update is the introduction of a dedicated “AI worker” for every Mac user, capable of assisting directly on the user’s computer. Even when users are manually operating their devices, Codex can work silently in the background without interrupting the use of other software.

According to Kavvy Lynch, OpenAI’s product management director, “Codex can operate applications on your computer in the background rather than taking over the entire computer’s operations.” This means that Mac users now have a dedicated AI assistant.

Developers often spend a significant amount of time coordinating communication and gathering information, leaving limited energy for actual programming. The upgrade aims to deeply integrate Codex into the operating system and developer tool ecosystem, freeing developers to focus on core tasks that require creativity and thought.

The new features of Codex are directly integrated into users’ existing ChatGPT accounts, requiring no additional API key configuration. Previously, users needed some programming knowledge to solve API integration issues, but now they can simply command Codex to generate the desired applications.

The updated Codex is now live, available to all users logging into the Codex desktop application via ChatGPT. The application supports both macOS and Windows systems, with specific features being gradually rolled out. The background computer control feature initially supports only macOS and will soon be available to users in the EU and the UK.

New Computer Use Feature: Codex Works in the Background

A key technology in this Codex update is the Computer Use feature, which allows Codex to break through the limitations of traditional chatbots by performing viewing, clicking, and input operations across all applications on your device. Importantly, all of this is done in the background.

This means Codex can now operate software as if it were a human watching the screen. It can understand, think, and execute operations. Moreover, multiple agents can run simultaneously on a Mac without affecting normal operations in other programs.

Windows users can still use and receive official support for the core Codex desktop application and extract information from Windows applications to display in Codex, but the initial rollout of the updated Codex does not support the same cursor-level background interaction feature as on macOS.

Unified Workspace: New Multi-Scene Features

In addition to system-level control capabilities, OpenAI has upgraded Codex into a unified workspace, covering the entire workflow from GitHub pull request reviews to remote infrastructure management. The updated Codex and its applications can now operate in a broader range of scenarios.

To cover the entire workflow of developers, the updated Codex has added several key features, including:

1. Built-in Browser: Developers can add comments directly on the in-app browser page, providing more precise instructions to the agents.
2. Visual Capabilities: By integrating gpt-image-1.5, along with screenshots and code, Codex can create product concepts, front-end designs, models, and visual effects for games within the same workflow.
3. Expanded Sidebar: The application now offers rich previews of non-code files such as PDFs, spreadsheets, and presentations, equipped with a summary panel to track agent plans and information sources.
4. Terminal and SSH: The updated Codex supports multiple terminal tabs and has launched an alpha testing feature for SSH connections to remote development environments.

Additionally, to connect these dispersed tasks, OpenAI has introduced over 90 new plugins for Codex, including development tools like Atlassian Rovo, CircleCI, and GitLab Issues. These plugins integrate skills, application integrations, and MCP servers, providing Codex with more ways to collect contextual information and execute related operations across tools.

VentureBeat quotes Ambrose as saying, “You can use @ mentions to specify applications for Codex to use; if not specified, Codex can also autonomously determine which applications to use.”

Support for Long-Term Task Execution and Enhanced Memory Function

OpenAI has also expanded Codex’s automation capabilities. It now supports reusing existing conversation threads while retaining built context. Codex can wake up automatically to continue executing long-term tasks that span days or even weeks, and it can autonomously schedule subsequent task operations.

In simple terms, Codex can now remember previous conversations and task progress. Unlike the previous method of “doing one thing at a time,” this update allows it to remember context and schedule its work. For example, when you go to sleep at night, it can automatically wake up to continue working.

Moreover, once Codex remembers your instructions, it can automatically schedule tasks across days or weeks, executing in the background and resuming from breakpoints, eliminating the need for you to repeat instructions daily until the task is completed.

For instance, product managers (PMs) previously needed to frequently switch between Slack, email, and Notion documents, manually syncing information with the development team, resulting in a heavy workload of copying and pasting.

With Codex’s automation capabilities, when a new requirement arrives in Slack, an email is received in Gmail, or a Notion document is updated, it can automatically capture and integrate relevant information into the development process, saving PMs from tedious application switching and enhancing their efficiency.

Following this logic, development teams can deploy the intelligent agents they desire, allowing Codex to manage various tasks automatically without manually opening and syncing each one.

It is worth noting that to enhance the efficiency of the aforementioned Codex automation features, OpenAI has also launched a preview version of the memory function.

Codex will remember key information from historical interactions, such as user preferences, historical corrections, and collected information, reducing the need for extensive customization instructions in each new session to improve processing efficiency.

In addition to executing your commands, Codex will also proactively suggest what to do next. By combining context, finding associated plugins, and recalling memory information, it can help you plan your workday or remind you where to continue from the last project.

For example, Codex can identify comments in Google Docs that need your attention, extract relevant information from Slack, Notion, and code repositories, and generate a prioritized to-do list for you.

Conclusion: Continuous Technological Upgrades and Expanding Human-Machine Collaboration

The comprehensive upgrade of Codex represents an extension of OpenAI’s capabilities in the AI development assistant field. By creating an agent logic for the Computer Use feature tailored for macOS, AI is evolving from simple conversational interactions to full-process automation and context-aware intelligent collaboration.

However, the new features introduced in this Codex update are still in the early preview stage, and the subsequent experience and ecosystem refinement will require ongoing observation.

Before this Codex update, Anthropic Claude had already launched a capability system directly comparable to Computer Use in its desktop products Claude Code and Claude Desktop in March, supporting out-of-the-box functionality for users to experience upon downloading the client.

As AI continues to penetrate operating systems and development toolchains, related technologies are generally moving towards more automated intelligent assistants. In this trend, the inefficient repetitive aspects of developers’ work will continue to be simplified, and the boundaries of human-machine collaboration may further expand in the future.