MCP vs WebMCP: Differences and Which to Use (2026)
MCP connects your AI assistant to services; WebMCP turns a website into tools for agents in the browser. Differences, examples and which fits you.
Nov 11, 2026· 4 min read
They have almost the same name and both let an AI use tools, but they work in different places. MCP connects your AI assistant to a service through a server. WebMCP makes a web page offer its tools to an agent that is visiting it in the browser.
The differences at a glance
| MCP | WebMCP | |
|---|---|---|
| What it is | A protocol to connect assistants to services | A browser API for a website to offer tools |
| Where it works | In your assistant: Claude, ChatGPT, Gemini CLI, VS Code… | Inside a page open in the browser |
| Who publishes it | The author of an MCP server | The website owner |
| What you install | You add the server to your assistant | Nothing: the tools live on the website itself |
| Sign-in | The server asks for its own authorisation (OAuth) | Uses the session you already have open on that site |
| Status | Standard in use, version 2026-07-28 | Draft, in testing in Chrome |
| Example | Context7 | asus.com |
What MCP is, in one line
MCP (Model Context Protocol) is the open standard used by Claude Fable 5.2, ChatGPT, Gemini CLI, Cursor and many others to connect to services and data. You install a server in your assistant and from then on the assistant can use its tools in any conversation. It's explained in depth in what MCP is and what it's for.
What WebMCP is, in one line
WebMCP is a proposal from the W3C's Web Machine Learning group. It lets a page tell the browser "these are my tools", with their name, description and the data they need. An agent visiting that page can use them directly instead of imitating clicks or reading the HTML blindly.
For developers it's a JavaScript function, document.modelContext.registerTool(), that registers each tool:
document.modelContext.registerTool({
name: "search_products",
description: "Search the catalogue by text and maximum price",
inputSchema: {
type: "object",
properties: {
text: { type: "string" },
maxPrice: { type: "number" }
},
required: ["text"]
},
async execute({ text, maxPrice }) {
return searchCatalogue(text, maxPrice)
}
})
The tool runs on the page itself, with the code the site already has, so it responds instantly and with the data the user sees.
One example of each
MCP: you connect Context7 to Claude once. From then on, in any chat you can ask "how do I do this in the latest version of this library?" and Claude fetches the current documentation through the server without you opening any website.
WebMCP: you visit asus.com with a browser that supports WebMCP. The site offers seven tools to the agent, such as searching products or checking the FAQs. You ask "show me gaming laptops under $1,500" and the agent uses the site's own search instead of clicking through buttons.
Current status of WebMCP (November 2026)
WebMCP is not stable yet:
- Chrome lists it as "proposed". Developers can try it since Chrome 146 by
turning on an experimental flag.
- There's an origin trial between Chrome 149 and 156, extended to 162, so
websites can test it with real users.
- The specification is a draft that changes every week. It uses
document.modelContext; some older tutorials say navigator.modelContext, which is the previous name.
- Registration through HTML attributes (without JavaScript), mentioned by some
websites, is not part of the specification.
Even so, there are already many websites with WebMCP tools in our WebMCP directory.
Which one do you need?
| Your case | What works for you |
|---|---|
| You want your assistant to look up data or use a service | MCP: find a server in the directory |
| You have a service and want it to work inside Claude or ChatGPT | MCP: publish an MCP server |
| You have a website with search, bookings or forms and want agents to use it well | WebMCP: start testing it |
| You want to reach both assistants and browser agents | Both |
For most users, today the answer is MCP: it's what already works in Claude, ChatGPT and Gemini. For website owners, WebMCP is what's coming, and testing it now is an advantage: browse WebMCP examples to see sites that already use it.
If you want to start with MCP, follow the guide on how to install an MCP.
Servers mentioned in this guide
Context7
CommunityContext7 pulls up-to-date, version-specific library documentation and code examples into AI coding agents via MCP.
GitHub
CommunityOfficial GitHub MCP server connecting AI assistants to repositories, issues, pull requests, Actions, security alerts, and more via natural language.
Frequently asked questions
Does WebMCP replace MCP?
No. MCP connects assistants to services from outside the browser; WebMCP works inside an open page. They complement each other.
Which browsers support WebMCP?
As of November 2026, only Chrome, in testing: for developers since Chrome 146 and as an origin trial between versions 149 and 162. It isn't switched on for everyone yet.
Do I need WebMCP for AI to understand my website?
Not yet. Today assistants read your site as text. WebMCP lets an agent do things on it (search, book, filter) reliably, and it's worth starting to test it.
Can I use WebMCP with Claude or ChatGPT?
Only through a browser or an agent extension that supports it. Regular assistants use MCP, not WebMCP.
Is WebMCP safe?
The tools run on the page itself, with the session and permissions you already have on that site, and the browser controls access. Still, it's a draft and security is still being discussed.
Which is easier for a company to set up?
If you already have a website with search or forms, WebMCP reuses the code you already have. If you want your service to work inside Claude or ChatGPT, you need an MCP server.
Browse the free directory of MCP servers, or tell us about yours.