WebMCP / comparison
WebMCP vs MCP
They share four letters and a good idea, so they get conflated constantly. But MCP is a client-server protocol that lives outside the browser, and WebMCP is a browser API that lives inside the page. Clearing that up is the fastest way to understand either one.
Why the names collide
MCP stands for the Model Context Protocol, an open protocol introduced by Anthropic in late 2024 for connecting AI models to external tools and data. It caught on fast, and "MCP" became shorthand for "let a model use tools." WebMCP arrived into that vocabulary, borrowed the tool concept, and put "Web" in front of it. The name signals a family resemblance, and it invites a wrong assumption: that WebMCP is simply MCP running on web pages. It is not. It is a separate mechanism, developed by a different group, that solves an adjacent problem in a different place.
Side by side
The same idea, two very different places.
Both expose tools to an AI. Where the tools live, how they are reached, and who runs them is what separates them.
| Dimension | MCP | WebMCP |
|---|---|---|
| Full name | Model Context Protocol | Web Model Context Protocol |
| What it is | A client-server protocol | A browser API |
| Where tools live | In a separate server process or endpoint | In the web page itself |
| How the AI reaches them | JSON-RPC messages over a transport (stdio or HTTP) | Direct calls in the browser via document.modelContext |
| Runs in the browser? | No, outside it | Yes, inside the page |
| Who exposes the tools | A developer or company running a server | A website, to whatever in-browser agent visits |
| Identity and auth | The server handles its own authentication | Inherits the browser session; tools act as the logged-in user |
| Origin | Anthropic, 2024; widely adopted across AI tooling | W3C Web ML Community Group draft, 2026; pre-standard |
| Maturity | Established and in production use | Origin trial; no mainstream consumer yet |
MCP, in one paragraph
MCP connects an AI application to capabilities that live outside it. An MCP host, the AI app, runs a client that opens a connection to an MCP server. The server exposes tools, along with resources and prompts, and the two sides exchange JSON-RPC messages over a transport, typically standard input/output for a local server or HTTP for a remote one. The server is where the work happens: it might wrap a database, a ticketing API, or a filesystem. The model never touches those systems directly; it asks the server, and the server answers. The whole arrangement sits on the infrastructure side, away from any web page a person is looking at.
WebMCP, in one paragraph
WebMCP puts the tools in the page. A website calls document.modelContext.registerTool() in its own JavaScript, or annotates a form in its own HTML, and an AI agent running in the browser reads that list and calls the tools directly. There is no separate server to run and no wire protocol between the page and the agent, because the agent is already inside the browser with the page. The tools run in the page's context, as the user who is signed in. It is a way for a website, not an infrastructure team, to speak to an agent that has arrived to act on a person's behalf.
How they fit together
They are layers, not rivals. Picture a travel company. Its back office runs MCP servers so an internal assistant can read the booking database and call the pricing API. Its public website exposes WebMCP tools so that when a customer's browser agent visits, it can search availability and start a booking without scraping the page. An advanced agent might use both in one task: MCP for the systems behind the business, WebMCP for the site in front of it. Understanding which one a conversation is about is most of the battle, and it is why they earn separate documentation rather than a single muddled explanation.
Which one is your question about?
- If you are connecting a model to a database, an internal API, or a back-end system, that is MCP. It is mature, and its specification is the place to start.
- If you are asking how your website should present itself to AI agents that browse it, that is WebMCP, and it is the subject of this reference. Continue to the imperative API or the declarative API, or read what readiness honestly means before you act.
Sources
- webmachinelearning/webmcp, the W3C Web Machine Learning Community Group repository.
- The Model Context Protocol specification and documentation (Anthropic), for MCP's client-server architecture and transports.
Reviewed 23 July 2026. WebMCP is pre-standard and changing; confirm API details against the primary source before building.
Keep going
You have the distinction. Now the mechanics.
See exactly how a page registers a tool, or turns a form it already has into one.