OpenAI Codex App Introduces In-App Browser with CDP & Visual Annotations
OpenAI's Codex app now features an in-app browser, enabling visual annotations, agentic browser navigation, and full Chrome DevTools Protocol (CDP) access.
The new in-app browser for OpenAI's Codex app bridges the gap between code generation and visual UI iteration.
The Evolution of Codex: From Text to Visual Iteration
Following the major April 2026 updates that transformed OpenAI's Codex into a full desktop control surface, developers are discovering one of its most powerful new features: the dedicated in-app browser. Designed to collapse the feedback loop between design and implementation, the Codex in-app browser gives you and the AI agent a shared view of rendered web pages right inside your conversational thread.
Rather than relying on text-heavy "prompt-and-pray" descriptions of layout issues, developers can now open local development servers (like http://localhost:3000) and let Codex literally "see" the application.
Point-and-Instruct: Annotation Mode
The standout capability of the in-app browser is Annotation mode. It fundamentally changes how frontend developers and designers provide styling and layout feedback to an AI coding assistant.
- Visual Targeting: You can select specific elements or hold
Shiftto select an entire visual area on the page. - Granular Styling Feedback: A configuration icon next to the text input lets you specify exact values for fonts, spacing, text, and color. You can preview these changes directly on the page before sending them to Codex.
- Contextual Fixes: By leaving a targeted comment like "This button overflows on mobile; wrap it without changing the card height," Codex knows exactly which DOM element to target in the source code.
Taking the Wheel: Agentic Browser Use
While manual annotations are great for user-guided fixes, Codex can also operate the browser autonomously. By installing the Browser plugin and calling @Browser, Codex can step into an active testing role.
Once invoked, the agent can click, type, and navigate through local previews. It can take screenshots, inspect rendered DOM states, download page assets, and run read-only JavaScript. Most importantly, it can verify that a CSS or component fix actually worked visually. Security is strictly enforced: Codex operates on an explicit allowlist/blocklist basis, ensuring the AI asks for permission before accessing or interacting with unverified URLs.
Developer Mode: Unlocking CDP Access
For advanced debugging and performance tuning, OpenAI has introduced Developer mode. By enabling "Enable full CDP access" in the browser settings, you grant Codex controlled access to the Chrome DevTools Protocol (CDP).
This allows Codex to act as an automated performance engineer. You can prompt Codex to capture a performance trace, inspect network traffic, examine applied styles, and diagnose memory leaks directly in the live browser. Because CDP access exposes sensitive browser internals, this feature requires explicit per-task approval to ensure data safety.
In-App Browser vs. Codex Chrome Extension
It is important to understand the boundaries of this tool. The in-app browser is strictly a sandboxed environment. It does not support authentication flows, cookies, existing tabs, or browser extensions. It is designed purely for local dev servers, file-backed UI components, and unauthenticated public pages.
For workflows requiring signed-in states (like navigating internal dashboards, authenticating via OAuth, or working inside enterprise SaaS apps), developers should use the dedicated Codex Chrome extension (invoked via @Chrome), which operates securely inside your standard logged-in browser profile.
Conclusion
The Codex in-app browser represents a significant leap forward for frontend and web development. By allowing AI to share your visual context, manipulate the DOM, and access DevTools, OpenAI is making UI iteration faster, more accurate, and far more intuitive.
Enjoyed this?
Get more posts like this delivered to your inbox.
Firecrawl Introduces Prometheus: A Forward-Deployed Agent for Web Data
Next →Xiaomi Open-Sources MiMo Code: A Free Terminal-Based Coding Agent with Persistent Memory
Related Articles
OpenAI Releases Codex Security: An Open-Source CLI and TypeScript SDK to Find, Validate, and Fix Security Vulnerabilities in Your Code
4 min read
Anthropic Releases MCP 2026-07-28 with Stateless Core and Extensions
5 min read