Codex Gains Computer Control, Browser, Plugins for Super App
OpenAI upgrades Codex with parallel agent computer use, in-app browser for web iteration, image generation, and 90+ plugins like Jira and Microsoft suite, converging on everything-app features currently MacOS-only.
Parallel Computer Use Enables Non-Interfering App Control
Codex now controls your Mac via multiple parallel agents that see, click, and type with a dedicated cursor, avoiding interference with your work. This mirrors Claude's computer use but executes faster. To test, prompt Codex to open Chrome incognito, navigate to google.com, then openai.com for latest model info—or launch Notes app and add text like "Codex computer use demo." Permissions prompt on first use, and agents handle apps without visible cursor movement from your view. Trade-off: Currently MacOS-only, with slow rollout and occasional bugs like failed updates.
In-App Browser Speeds Web and Frontend Iteration
Built-in browser renders pages for direct feedback loops: inspect elements, add comments like "Add ability for user to select specific folder before indexing," then regenerate code. Uses existing GPT-4o model—no new model. Pinpointing UI elements provides precise context, outperforming link-based prompts. Ideal for frontend/game dev; future expansions could control all desktop apps. Demo showed adding folder selection path to a speech-to-text indexing UI, though click-to-select paths remain a wishlist item.
Image Gen and 90+ Plugins Boost Productivity Beyond Code
Integrates OpenAI image gen akin to Google's Imagen for UI ideation, plus 90 plugins combining apps, integrations, and MCP servers. Day-one options include Jira, CircleCI, GitLab issues, Microsoft suite, Remotion Render—useful for devs and non-devs in knowledge work. Enables context gathering/actions across tools without leaving Codex. Opinion: Refocuses OpenAI on coding/super-app strengths, ditching resource-draining side quests; UI redesign beats rushed Claude desktop. Rivalry with Anthropic accelerates features, but differentiation shrinks to execution quality.