====== WebMCP ======

**WebMCP** (Web Model Context Protocol) is a browser-native implementation of the Model Context Protocol (MCP) that allows websites to expose structured tools and data directly to AI agents through the ''navigator.modelContext'' JavaScript API.((Google and Microsoft, W3C Community Group. [[https://webmachinelearning.github.io/webmcp/|WebMCP Specification]])) First published in August 2025, WebMCP enables web applications to function as client-side MCP servers, eliminating the need for separate backend infrastructure to expose application capabilities to AI agents.((Brandon Walderman et al., "WebMCP: Enabling web apps to provide JavaScript-based tools for AI agents," August 2025. [[https://github.com/webmachinelearning/webmcp|GitHub Repository]]))

===== Background =====

The Model Context Protocol, introduced by Anthropic in November 2024, established a standard for AI agents to communicate with external tools and data sources via server-side JSON-RPC connections.((Anthropic, "Model Context Protocol," November 2024. [[https://www.anthropic.com/news/model-context-protocol|anthropic.com]])) While MCP proved effective for backend services, databases, and headless APIs, it left a gap for browser-based applications where business logic, user state, and authenticated context already resided in the client.((Zuplo, "WebMCP: How Websites Will Expose Tools to AI Agents," March 2026. [[https://zuplo.com/blog/what-is-webmcp|zuplo.com]]))

WebMCP addresses this gap by bringing MCP principles into the browser environment through a W3C Community Group standard developed jointly by Google and Microsoft.((DiscoveredLabs, "Google WebMCP Explained," 2026. [[https://discoveredlabs.com/blog/google-webmcp-explained-the-complete-guide-to-web-model-context-protocol-for-ai-agents|discoveredlabs.com]]))

===== Architecture =====

WebMCP defines two primary types of data that websites expose to agents:

  * **Context** — all data agents need to understand user activity, including content not currently visible on screen
  * **Tools** — actions agents can perform on behalf of users, from answering questions to completing forms and executing transactions

Unlike server-side MCP, which uses persistent server connections over HTTP or stdio, WebMCP operates entirely client-side. Web developers register structured tools via the ''navigator.modelContext'' API, and browser-based AI agents can discover and invoke these tools directly without intermediate infrastructure.((Chrome for Developers, "When to use WebMCP and MCP," March 2026. [[https://developer.chrome.com/blog/webmcp-mcp-usage|developer.chrome.com]]))

==== MCP vs. WebMCP ====

The Chrome team uses the analogy of a customer service call center versus an in-store expert:((Chrome for Developers, "When to use WebMCP and MCP," March 2026. [[https://developer.chrome.com/blog/webmcp-mcp-usage|developer.chrome.com]]))

  * **MCP** is available on any platform at any time, pulling up data and handling core tasks as needed
  * **WebMCP** is available only on a specific website, helping agents understand the UI and interact with application-specific capabilities

Both share the same philosophy of structured tool discovery and predictable execution, but they address different deployment contexts. They are complementary, not competing, standards.

===== Key Features =====

  * **Structured tool discovery** — agents can programmatically ask "what can you do?" and receive a machine-readable list of available tools, parameters, and descriptions
  * **Predictable execution** — explicit function calls replace guesswork, producing predictable outcomes
  * **No server required** — the website itself becomes the tool surface, reusing existing application logic
  * **Shared context** — agents and users work within the same web interface, maintaining shared state and user control
  * **Human-in-the-loop** — designed for collaborative workflows where users retain oversight of agent actions

===== Development and Adoption =====

The WebMCP specification was authored by researchers from Microsoft (Brandon Walderman, Leo Lee, Andrew Nolan) and Google (David Bokan, Khushal Sagar, Hannah Van Opstal).((WebMCP GitHub repository. [[https://github.com/webmachinelearning/webmcp|github.com/webmachinelearning/webmcp]])) The repository has attracted over 2,100 stars on GitHub and 20 contributors as of early 2026.

In February 2026, Google announced that WebMCP was available for its Early Preview Program (EPP), allowing developers to experiment with the API in Chrome.((DEV Community, "WebMCP Is Available for Early Preview: What You Need to Know," March 2026. [[https://dev.to/onsen/webmcp-is-available-for-early-preview-what-you-need-to-know-3b32|dev.to]]))

===== See Also =====

  * [[model_context_protocol|Model Context Protocol (MCP)]]
  * [[ai_agents|AI Agents]]

===== References =====