v149 · Origin Trial · AI Agents

WebMCP

A browser-native implementation of the Model Context Protocol — sites register structured tools that AI agents can discover and invoke directly, without scraping HTML or relying on hidden APIs.

Origin Trial (Chrome 149–156). WebMCP requires an origin trial token to enable navigator.modelContext. These demos simulate the API surface with mock implementations so you can explore the concept before registering for the trial.

concepts

  1. Tool Registry Explorer

    Register MCP tools on a page using the navigator.modelContext API, then inspect what an AI agent would discover. Define tool schemas, input parameters, and handler functions — live, in the browser.

  2. Agent Handshake Simulator

    Watch the full agent–page interaction unfold: a simulated AI agent discovers your registered tools, selects the right one for a task, sends a structured call, and receives a response. Step through each phase with annotations.

  3. Tool Consent Panel

    Simulate how a browser might surface MCP tool consent to users. Register a set of tools, inspect their JSON schemas, then grant or revoke permission per tool — exactly the control surface a real agent controller would expose.

  4. Schema Validator

    Paste or edit a JSON tool definition and get instant feedback: required fields, naming conventions, inputSchema completeness, and a preview of exactly what an AI agent will see when it discovers your tool. Includes three example schemas to load and explore.

  5. Multi-Tool Orchestrator

    Three complete agentic scenarios — shopping assistant, travel planner, code analyser — each using a chain of 3 tools. Watch the agent plan, call each tool in sequence, and synthesise a final answer. Animated step-by-step trace with call/response payloads.

  6. Error Handling Lab

    Six error scenarios a WebMCP tool call can produce — validation error, runtime error, timeout, permission denied, partial success, and tool not found — each with a full animated call/response flow, the structured error schema, and the recommended recovery strategy. A live simulation runs each scenario step-by-step so you can see exactly what the agent, browser, and site exchange.

    Simulation Error handling Recovery

why it shipped

AI agents interacting with the web today face a fundamental mismatch: websites are designed for human eyes — HTML, visual layouts, implicit affordances. Agents scrape, guess at selectors, and break on every redesign. The Model Context Protocol (MCP) addresses this by letting sites expose machine-readable capability descriptions: here is what I can do, here is how to call it, here is the schema of the result. WebMCP brings MCP natively into the browser via navigator.modelContext, so any installed AI (Gemini in Chrome, a local LLM, an agentic extension) can discover and invoke tools without HTTP calls to a separate server, without scraping, and with full user-consent visibility. The origin trial starting in Chrome 149 lets developers experiment with the API and give feedback before it becomes a baseline feature.

references

implementation reference

Need the exact API surface, compatibility boundaries, errors, lifecycle, and source links? Read the matching gendn reference ↗