Skip to content

OpenAI

Websites can hand Codex their own tools

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: OpenAI models and products →
Video: twitter.com

ChatGPT and Codex can now call the actions a website declares for agents. This works in the ChatGPT desktop app’s built-in browser and in ChatGPT Sites, OpenAI said in an August 26 post from OpenAIDevs on Twitter. When either one opens a compatible site, it can use that site to finish your task.

The standard behind it, WebMCP, is still experimental, by OpenAI’s own description. OpenAI calls it an open standard that lets websites expose structured tools agents can use directly, offered alongside the interface people already click through.

The site does the setup. The draft specification describes WebMCP as a JavaScript interface for web developers. A developer registers “tools”: JavaScript functions, each with a natural-language description and a structured schema for its inputs. Agents can invoke them, and so can a browser’s own agent or assistive technologies. In effect, the page works as a Model Context Protocol server whose tools live in client-side script instead of on a backend.

The tools run in the page itself and reuse the application logic the site already has, according to the spec. The user and the agent work in the same web interface and share context, and the user keeps control.

An earlier update gave the built-in browser URL suggestions as you type. With the WebMCP release, OpenAI also opened a WebMCP Challenge, a 10-day window for exploring what the standard can do.

To try it, update the ChatGPT desktop app to the latest version, open a site that supports WebMCP, and ask Codex to do the task there.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.