What is the Codex integration?
Codex can follow a Remote Browser skill to perform browser work using the tools available in its environment. Remote Browser supplies the separately running Chromium session, while the skill describes connection handling, session rules, and verification steps.
Official OpenAI documentation describes skills as reusable workflows with a SKILL.md file and optional supporting resources. The browser runtime and its endpoint are supplied separately; a skill is not a browser service by itself.
Why give Codex a remote browser?
A repository change can require a check against a running application, an authenticated account, or a browser environment different from the coding worker. A remote session gives the workflow a concrete runtime that a human can observe when needed.
For example, define a UI verification task with the application URL, expected account, steps, and visible completion condition. This makes the difference between code inspection and a verified browser result clear. Keep screenshots and task reports free of connection credentials.
Quick start
Use the Remote Browser Agent skill setup to install the product workflow and configure its endpoint. Confirm that the skill is discoverable in your Codex environment and that the required browser tools are available.
Set REMOTE_BROWSER_CDP_URL securely for the process that will make the connection. A first task can be explicit about the intended skill and result:
Use the Remote Browser Agent skill with the configured endpoint.
Open https://example.com and verify the page title and final URL.
Do not print the endpoint. Report only the result you observed.This prompt is an example verification task, not a new command syntax. Check the viewer or returned browser evidence before expanding the workflow to an authenticated application.
Connection methods and runtime boundaries
The skill route follows the tools and conventions described in the installed skill. An MCP route uses a separately configured browser server; the Playwright MCP collection explains that architecture. Both need the intended browser endpoint and a clear owner for navigation.
Codex's workspace and the remote browser's filesystem are separate boundaries. A local build on a worker is not automatically reachable from the remote browser. Use an authorized reachable preview or deployment and test the actual route from the browser environment.
Use cases
- Check a deployed UI after implementing a repository change.
- Reproduce a login-dependent bug with an authorized test profile.
- Inspect browser behavior while a human observes the session.
- Capture the visible result needed to assess an acceptance criterion.
Make the account, target URL, and allowed actions part of the task. If a connection drops during a submission, inspect the resulting state before retrying it.
FAQ
Does installing a skill automatically provision a browser?
The skill instructions and the runtime allocation are separate. Follow Remote Browser's setup to supply the endpoint and verify tool access.
Where should I check skill discovery?
Use the official OpenAI documentation for the current Codex environment and compare it with the installer's documented destination. Avoid assuming that every client uses the same directory layout.
Is browser attachment time a Codex task benchmark?
No. A task benchmark also includes agent execution, tool calls, application behavior, and verification. Use the CDP timing method as one component and report the remaining phases separately.