OpenAI Upgrades Codex Into a Full Desktop Agent With Computer Use, 111 Plugins, and Parallel Workflows
OpenAI's April 16 Codex update adds background computer use on Mac, parallel multi-agent execution, persistent memory, and 111 plugin integrations, escalating the agentic coding tool race with Anthropic.
Overview
OpenAI released a major expansion of its Codex desktop application on April 16, 2026, rebranding the update under the banner “Codex for (almost) everything.” The release adds background computer use on macOS, the ability to run multiple agents simultaneously without interfering with a user’s active workflow, a persistent memory system, an in-app browser, and a catalog of 111 new plugin integrations — a direct competitive response to Anthropic’s Claude Code, which OpenAI acknowledged has become widely adopted in enterprise developer environments.
What Changed
The headline capability is background computer use: Codex can now operate macOS applications with its own cursor, viewing screen content and executing clicks and keyboard input while the user continues working in other apps. According to TechCrunch, this allows “multiple agents” to run “in parallel, without interfering with your own work in other apps” — a design specifically aimed at long-running development tasks such as test suite execution, frontend iteration, and GUI-only bug reproduction.
The update also introduces thread automations, a feature that allows Codex to schedule future work and resume long-running tasks across days or weeks without requiring the user to restart a session. A new artifact viewer supports previewing PDFs, spreadsheets, and presentations directly inside the app. An in-app browser lets users annotate rendered pages with comments, feeding precise instructions to the agent for frontend and game development tasks, as OpenAI describes. Image generation using GPT-image-1.5 is now embedded in the workflow, enabling the creation of mockups, slide visuals, and placeholder assets without leaving the development environment.
On the integration side, TechCrunch reports that OpenAI added 111 plugin integrations — higher than earlier reports of “more than 90” — including CodeRabbit, GitLab Issues, Atlassian Rovo, CircleCI, Render, and Microsoft Suite connectors. The company also added a GitHub pull request review workflow inside the sidebar, the ability to address review comments directly in the agent, and multiple terminal tabs for simultaneous shell processes.
Changelog Details
The official Codex changelog documents the version 26.415 release with additional specifics: remote SSH connections entered alpha, enabling Codex to operate against remote development boxes rather than only local machines. A task sidebar now surfaces plans, sources, and summaries. “Chats” mode was introduced to allow users to start threads without first selecting a project folder, lowering the friction for ad-hoc queries. A CLI companion release, version 0.122.0 (published April 20), added filesystem deny-read glob policies, isolated execution modes, and tool discovery enabled by default.
Competitive Context
The update arrives three weeks after Anthropic brought computer use to macOS through Claude Code and Claude Cowork, as previously reported. TechCrunch noted that Claude Code has been “dubbed the tool of choice for many businesses,” with Anthropic’s tool gaining particular traction for handling large, complex codebases. Codex has historically been characterized as faster in execution but weaker in sustained, multi-step workflows — the gap this update directly targets.
The Microsoft Agent Framework 1.0, released April 7 and covered by The Machine Herald, explicitly listed the Codex SDK as a composable agent harness within multi-agent pipelines, signaling that the broader developer ecosystem already treats Codex as infrastructure rather than just a standalone tool.
Availability and Limitations
The computer use features are limited to macOS at launch. Enterprise, Education, EU, and UK users will receive memory and personalization features in a subsequent rollout phase, according to the Codex changelog. Remote SSH devboxes remain in alpha. A new pay-as-you-go pricing tier was introduced for ChatGPT enterprise and business customers, according to TechCrunch.
What Remains Unclear
OpenAI has not published independent benchmark comparisons between the updated Codex and competing tools for multi-step agentic coding tasks. The security implications of an agent with persistent background computer use, cross-application access, and scheduled autonomy have been flagged by security observers but not formally addressed in OpenAI’s documentation. The degree to which the 111 plugins are available to all tiers versus enterprise-only customers is not specified in the release materials.