WebMCP lets web apps expose standardized tool interfaces for AI agents alongside their human UI, built on open web standards via a W3C community group.
Adapted from @dkundel# WebMCP: building for agents on the open web WebMCP lets you build web apps that you and your agent can use together. It builds on the open web: your website exposes tools alongside its UI, and your agent discovers them as it uses the site. We recently added WebMCP support to the in-app browser in the ChatGPT desktop app. WebMCP is an experimental proposal for a web standard. It was created by Google and Microsoft and is being developed through a W3C community group. (https://webmachinelearning.github.io/webmcp/) # Give your agent its own interface With computer use, an agent clicks around through the same interface you provide to a person. Even if you use Playwright, the application is still designed for people first. You could hide metadata in the application, but the agent needs to know to look for it. WebMCP gives the agent a standardized interface to the tools your application exposes. You can have one interface for the user and a different set of actions for the agent, without showing both to the user. For something like 3D modeling software, you can expose actions that make sense for an agent. It doesn't have to click all the same buttons you designed for a person. At the same time, you can expose data to the agent without showing it in the UI. That could be helpful metadata that the agent can use but the user doesn't need to see. # Why don't you just use MCP then? The first difference is discovery. With regular MCP, you need to know an integration exists, when it's needed, and when to set it up. It's similar to the difference between a mobile app and a mobile website. An app has its own life cycle and may go through a store. For example, ChatGPT has curated plugins, but you have to know when to use them. With WebMCP, your agent can discover tools when it visits a website on your behalf. The tools are exposed by the page you're already using, so you don't need to find and connect a separate MCP integration first. A listing in a well-known directory could help an agent discover an MCP, but it doesn't by itself remove that setup step or put the agent in your current browser session. # Sharing the same browser session The second difference is collaboration. With WebMCP, you can share the same browser session with your agent without needing a fancy backend. A good example is my Margin Editor that I built. You and the agent can both leave comments, and the UI treats them differently. You could also leave a comment for the agent without displaying it to every other user. (https://margin-local-docs.openai.chatgpt.site/) You could store the document entirely in local storage. Your agent interacts through WebMCP while you edit through the UI. You're working on the same document in the same browser session, without routing everything through a server backend or a local server. Imagine a game that you're cooperatively playing with the agent in that same browser session. # How do you use WebMCP? When ChatGPT or Codex opens a website in the in-app browser or cloud browser, it can discover the WebMCP tools that site exposes and use them when relevant. To see which tools a site exposes, click the cursor icon in the in-app browser's URL bar and expand Available site tools. To try WebMCP locally in Chrome while it's experimental, open chrome://flags/#enable-webmcp-testing, set the flag to Enabled, and relaunch Chrome. You can find the steps in Chrome's WebMCP documentation. (https://developer.chrome.com/docs/ai/webmcp#local-webmcp) # How to build with WebMCP The best way to get started with WebMCP is to just ask Codex to add it to the website or app that you are building. If you want to add it yourself, check whether the browser supports WebMCP, then define tools in JavaScript the same way you'd define functions for most model APIs. This is a short example, and the spec is still evolving. Check out the WebMCP JavaScript API for the details. (https://webmachinelearning.github.io/webmcp/) # WebMCP Challenge If you feel inspired and want to explore what an agent native web with WebMCP might look like, join our WebMCP Challenge. Build something new or add WebMCP to an app you already have. Submissions close on September 3 at 1 p.m. PT. (https://openai.com/webmcp-challenge/) Join us for Discord office hours on August 31 at 11 a.m. PT. Bring your questions and get help with your project. (https://discord.com/channels/974519864045756446/1517482456964534312) I’m excited to see what you’ll create for an agent native open web.
