WebMCP: The Web Is Learning to Talk to AI Agents

Posted on Edited on In , Symbols count in article: 5.7k Reading time ≈ 10 mins.

The web was built for humans — click buttons, fill forms, scroll pages. AI agents trying to use these same interfaces resort to brittle hacks: scraping DOM trees, parsing screenshots, or simulating mouse clicks on <div>s pretending to be buttons. It works, barely, and breaks constantly.

WebMCP changes this. It lets web developers explicitly declare what their site can do for AI agents — turning every web page into an MCP server that runs client-side.

What Is WebMCP?

WebMCP is a joint proposal from Microsoft Edge and Google Chrome teams (published August 2025) that adds a JavaScript API to the web platform. It lets developers register “tools” — functions with natural language descriptions and structured schemas — that AI agents can discover and invoke directly.

Think of it as: MCP, but the server lives in your browser tab.

Here’s a minimal example — a TODO app exposing an “add task” tool:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
window.agent.provideContext({
  tools: [
    {
      name: "add-todo",
      description: "Add a new todo item to the list",
      inputSchema: {
        type: "object",
        properties: {
          text: { type: "string", description: "The text of the todo item" }
        },
        required: ["text"]
      },
      async execute({ text }) {
        // Your existing app logic — same code the UI uses
        addTodoItem(text);
        return { success: true, id: newItem.id };
      }
    }
  ]
});

The agent doesn’t need to find a text field, type into it, and click “Add.” It calls add-todo directly, using the app’s own code path.

Why This Matters

1. Developer Control Over Agent Behavior

Today, agents interact with your site however they can — often poorly. WebMCP puts you in the driver’s seat: you decide what actions are available, what they’re called, what parameters they accept, and what they do. This is consent-based agent interaction, not adversarial scraping.

2. Shared Context Between Human and Agent

Unlike backend MCP servers that operate in isolation, WebMCP tools run inside the page the user is looking at. The agent sees the same state the user sees. This enables cooperative workflows:

“Find me summer wedding dresses, red or orange, no sleeves” → agent filters options using WebMCP tools → user browses the results visually → user picks one → agent handles checkout.

Human and agent, same interface, shared context. No context-switching between chat windows and web pages.

3. Authentication for Free

This is the killer feature. Backend MCP servers need OAuth flows, API keys, token management. WebMCP tools execute in the browser tab where the user is already logged in. Cookies, JWTs, session tokens — all just work. Zero auth configuration.

4. Accessibility as a Bonus

The same challenges AI agents face — understanding page structure, finding actionable elements, dealing with custom components — are the same challenges assistive technologies have battled for decades. WebMCP tools benefit screen readers and other AT just as much as AI agents.

The Landscape: Three Competing Approaches

The “MCP in the browser” space is surprisingly crowded already:

Approach How It Works Status
W3C WebMCP (Microsoft + Google) Native browser API, window.agent.provideContext() W3C proposal, Edge prototype
MCP-B (community) Chrome extension bridges tab MCP servers to external clients Open source, working today
Jason McGhee’s WebMCP JS library, blue widget, any site can add it webmcp.dev, independent project

The W3C proposal is the most ambitious — it aims to become a native browser API, meaning no extensions needed. MCP-B is the most practical today — it works via a Chrome extension that bridges tab-level MCP servers to clients like Claude Desktop or Cursor. McGhee’s original WebMCP was the proof-of-concept that validated the idea, leading to the W3C standardization effort.

The Implications

Agent SEO Is Coming

If agents can discover which sites offer which tools, then optimizing for agent discoverability becomes a thing. The WebMCP spec explicitly calls out that “tool discoverability is not yet a solved problem.” Expect a new discipline — Agent SEO — to emerge alongside traditional search optimization.

Two-Layer Websites

Sites will increasingly have:

  • Human layer: Visual, branded, narrative — what you see today
  • Agent layer: Structured, schema-based, fast — WebMCP tools

Some operations will go entirely through agents (reordering supplies, scheduling appointments), with the human UI only appearing for approval steps or edge cases.

The Attention Economy Gets Disrupted

If an agent can book a flight by calling search-flights and book-ticket directly, it never sees your homepage banner ads. The advertising model that funds the free web doesn’t have an obvious answer to this yet.

My Take

As someone running AI agents daily (I literally have agents managing my calendar, monitoring my inbox, and posting to my blog), WebMCP solves a real pain point. Current browser automation — whether through Playwright, puppeteer, or vision-based agents — is slow, expensive, and fragile. A 10-step form fill that takes 20 seconds and costs $4 in API calls could become a single tool call that executes in milliseconds.

The biggest question isn’t technical — it’s adoption. WebMCP only works if websites implement it. And websites only implement it if there’s a business incentive. The W3C backing from both Google and Microsoft is encouraging, but we’ve seen promising web standards stall before.

The most likely path: major platforms (Google Flights, Amazon, banking sites) adopt first because they already have backend APIs — wrapping them as WebMCP tools is trivial. Then frameworks (Next.js, Shopify, WordPress) build it in as a feature. Then it becomes default.

We’re probably 2-3 years from WebMCP being ubiquitous. But the direction is clear: the web is learning to talk to AI agents, and developers get to control the conversation.

References

  1. W3C WebMCP Proposal — Microsoft Edge + Google Chrome teams (Aug 2025)
    github.com/webmachinelearning/webmcp

  2. “AI agents and the web: A proposal to keep developers in the loop” — Patrick Brosset, Microsoft Edge (Aug 2025)
    patrickbrosset.com/articles/2025-08-28-ai-agents-and-the-web

  3. MCP-B: Model Context Protocol for the Browser — Community project
    mcp-b.ai

  4. WebMCP (Jason McGhee) — Original independent implementation
    webmcp.dev

  5. Model Context Protocol Specification — Anthropic (Nov 2025)
    modelcontextprotocol.io/specification/2025-11-25

  6. “What is WebMCP? Agentic Web and Websites Interacting with AI Agents” — Zeo
    zeo.org/resources/blog/what-is-webmcp

  7. “The Model Context Protocol’s impact on 2025” — ThoughtWorks (Dec 2025)
    thoughtworks.com/insights/blog/generative-ai/model-context-protocol-mcp-impact-2025