OpenAI shipped a major update to Codex on April 16 that turns its coding agent into something closer to a general-purpose desktop agent. Codex can now operate macOS desktop apps with its own cursor, run multiple agents simultaneously in the background, schedule its own future work, generate images, and browse the web through a native in-app browser.
The update shipped to Codex desktop users signed in with ChatGPT, initially limited to macOS. EU and UK users will have to wait for computer use and personalization features, according to OpenAI’s blog post.
Computer Use on macOS
The headline feature: Codex can now see, click, and type in any macOS desktop app using its own cursor. It operates in the background, meaning multiple agents can work across different apps without taking over the user’s screen or interrupting their workflow.
“For developers, this is helpful for testing and iterating on frontend changes, testing apps, or working in apps that don’t expose an API,” OpenAI wrote in its announcement.
The computer use capability came from OpenAI’s acquisition of Sky Applications Incorporated, the team behind Apple Shortcuts, according to 9to5Mac. Ari Weinstein, who led that team, posted on X that the feature lets agents use Mac apps concurrently while the user continues their own work.
Autonomous Scheduling and Memory
Codex can now schedule future work for itself and wake up automatically to continue long-running tasks, potentially across days or weeks. It can also resume work using existing conversation threads, preserving context built up in prior sessions.
A new memory feature, launching as a preview, lets Codex remember preferences, corrections, recurring workflows, and information from past sessions. Using that stored context plus data from connected plugins and projects, Codex can suggest how to start the work day or where to pick up on a previous project. MacRumors confirmed the memory feature covers tech stacks, personal preferences, and user-specific workflow patterns.
Browser, Image Generation, and Plugins
An in-app browser built on OpenAI’s Atlas lets users comment directly on web pages to give the agent precise visual instructions. OpenAI plans to expand this so Codex can fully command the browser for navigating websites, testing user flows, and inspecting outputs.
Image generation via gpt-image-1.5 is now integrated directly in Codex, removing the need to switch to ChatGPT for visual mockups and product concepts. OpenAI also shipped 111 new plugins combining skills, app integrations, and MCP servers, along with support for multiple terminal tabs, addressing GitHub review comments inline, SSH connections to remote devboxes (in alpha), and rich file previews for PDFs, spreadsheets, and slides.
New plugins include integrations for GitLab, Atlassian Rovo, and Microsoft Suite, connecting Codex to the enterprise developer toolchain beyond just code repositories, according to The Verge.
The Competitive Context
The Verge characterized the update as “a direct shot at Claude Code,” noting it arrives as OpenAI’s rivalry with Anthropic intensifies following Claude Code’s rapid adoption and OpenAI “aggressively shifting resources” to compete.
Codex now has 3 million weekly users, a 5x increase in three months with 70% month-over-month growth, according to 9to5Mac. OpenAI also recently launched a $100/month Pro subscription tier aimed at heavy Codex users, offering 10x usage limits compared to the $20/month Plus plan.
The Desktop Agent Dimension
The autonomous scheduling capability is the most architecturally significant piece. A coding agent that can set its own wake-up timers, resume work across multi-day horizons, and propose next actions based on accumulated memory is no longer just a tool you invoke. It is a persistent agent that operates on its own timeline.
Combined with computer use (controlling any macOS app, not just code editors), Codex has crossed from “AI coding assistant” into “desktop agent that happens to be very good at code.” The question for teams evaluating AI coding agents: whether that breadth introduces more risk surface than it resolves in productivity.