ChatGPT Site Tools Bring WebMCP Actions to the Desktop Browser

ChatGPT can use structured tools exposed by websites

OpenAI has documented site tools for the ChatGPT desktop app, an agent-browser capability built on WebMCP that lets compatible websites expose structured actions directly to ChatGPT.

Instead of relying only on visual browsing, clicks and typing, ChatGPT can discover tools a webpage provides and use them to search, edit or carry out supported tasks while the user remains on the same live page. OpenAI's developer announcement for agent-ready websites appeared on August 25, and its Help Center documentation was updated again on August 28.

This is a released capability with gated availability, not a finalized web standard. OpenAI says site tools are available only when the user's account and selected model support them and when the open webpage actually exposes a compatible tool.

WebMCP is the underlying proposed standard

WebMCP is an experimental/proposed web standard for exposing browser-page functionality as structured tools that AI agents can discover and invoke. The Web Machine Learning Community Group's specification allows websites to register tools with names, descriptions and input schemas instead of forcing an agent to infer every action from pixels or the DOM.

OpenAI's site tools are an implementation of this broader WebMCP approach. They operate in the ChatGPT desktop app's built-in browser. OpenAI says no separate connector is required for a supported page, but the feature is currently not the same as using ordinary Chrome browsing.

The live page and signed-in session stay in context

A notable design choice is that site tools work with the webpage the user already has open, including its current state and signed-in session. That can reduce the need for a developer to recreate browser state through a separate backend integration for some interactive workflows.

Potential uses listed by OpenAI include searching documentation, editing documents or shared workspaces, exploring dashboards, building travel itineraries and updating shopping carts. Each website decides which tools to expose, and tools are scoped to the page that provides them rather than automatically carrying across unrelated sites.

OpenAI adds safety checks around tool use

OpenAI warns that site tools introduce risks including data exfiltration and prompt injection. Its documentation says ChatGPT asks for website permission and performs additional checks before tool calls. Sensitive actions such as purchases, deletion, permission changes, sending messages or sharing personal information require confirmation.

The documentation also says instructions from a website or site tool cannot themselves authorize ChatGPT to disclose information or perform sensitive actions on the user's behalf.

These protections reduce risk but should not be read as eliminating browser-agent security problems. Developers still need narrowly scoped tools, robust validation, clear descriptions and safe authorization boundaries.

What is available now versus still evolving

Available now, for eligible accounts and compatible websites, is ChatGPT's ability to discover and use site tools inside the desktop app's built-in browser. OpenAI also provides a WebMCP Challenge where developers can test agent-native applications.

Still evolving is WebMCP itself. It remains a proposed/experimental standard rather than a final universally supported browser specification, and implementations can change as the standard develops.

For developers, the important shift is architectural: websites can begin exposing intentional, typed actions to browser agents instead of depending entirely on screen interpretation and simulated human input. Whether WebMCP becomes broadly interoperable will depend on standards evolution, browser adoption, security design and real-world developer uptake.