Feature request: First-class Chrome integration with DOM+vision understanding

Summary

As a heavy Codex user and web developer, I want a first-class Chrome integration similar to Claude Code's Chrome beta (https://code.claude.com/docs/en/chrome). The goal is a smooth, reliable workflow where Codex can drive a real browser, understand the HTML DOM, and fuse DOM context with screenshots for accurate UI reasoning.

Problem

Today, using browser tooling via MCP/Playwright/DevTools is brittle and manual. There is no unified, first-class Chrome integration in Codex that feels smooth for day-to-day web development. The agent often lacks a dependable way to correlate what it sees in a screenshot with the underlying DOM or computed layout, which limits debugging and UI automation.

Desired capabilities

• First-class Chrome connection: connect to an existing Chrome or launch a managed instance; auto-reconnect; stable long sessions.
• DOM-aware page understanding: structured DOM snapshot, computed styles, layout boxes, accessibility tree.
• Vision + DOM alignment: map screenshot coordinates to DOM nodes and vice versa; highlight bounding boxes; select elements by clicking or describing what is visible; screenshot overlays for debugging.
• Built-in DevTools signals: console logs, network requests, performance traces, storage (cookies/localStorage) accessible to the agent.
• Interaction primitives: click, type, scroll, hover, select, file upload, keyboard shortcuts.
• Smooth UX: one command to enable; low latency; reliable session management; works in Codex CLI sandbox; cross-platform (macOS, Windows, Linux, WSL).
• Security controls: explicit user approval for page access and data exfiltration; domain allowlist; safe-mode.

Use cases

• Visual debugging: identify layout issues, CSS conflicts, or hidden elements by linking screenshot anomalies to DOM/CSS.
• Frontend iteration: change UI and verify immediately using screenshots + DOM validation.
• QA and automation: reproduce bugs, run checks, and validate flows end-to-end.

Acceptance criteria

• From Codex CLI, a user can enable Chrome integration in one step, list open tabs, and open a new tab.
• For a typical debugging task, Codex can take full ownership: open the target page, reproduce the issue, inspect DOM/CSS/console/network, and return a diagnosis with evidence (selectors, logs, screenshots) without manual browser操作.
• The agent can answer "what element is this in the screenshot?" with a stable DOM selector, bounding box, and an optional highlight overlay.
• The agent can select elements by visual description or direct click, then confirm via DOM + screenshot alignment.
• All common browser operations (click/type/scroll/hover/select/upload/keyboard shortcuts) are exposed as reliable primitives and can be chained end-to-end by Codex.
• The integration is stable for long sessions, recovers from stale browser state automatically, and works across macOS/Windows/Linux/WSL in the Codex CLI sandbox.

Related

• Claude Code Chrome docs: https://code.claude.com/docs/en/chrome
• This issue is a feature request for a first-class, smooth Chrome integration, beyond the current ad hoc MCP setup.