v149 · WebMCP · agentic AI

Multi-Tool Orchestrator

AI agents rarely call a single tool. Real tasks chain several: search → filter → transform → present. This simulator shows how an agent walks through a registered WebMCP tool chain step by step, with full call/response traces for each hop.

Origin Trial (Chrome 149–156): navigator.modelContext requires an OT token. This demo simulates tool calls with mock data — no token or microphone needed.

Pick a scenario, then watch the agent chain the tools:

Agent execution trace
Pick a scenario and click Run agent.
// An agent receives a user task and decides which tools to call
async function runAgentTask(task, tools) {
  const plan = await llm.plan(task, tools); // e.g. ["search_products","filter_results","format_cart"]

  let context = {};
  for (const toolName of plan) {
    const tool = tools.find(t => t.name === toolName);
    const args = await llm.buildArgs(tool.inputSchema, context, task);
    // WebMCP: the browser routes the call to the page's registered handler
    const result = await navigator.modelContext.callTool(toolName, args);
    context[toolName] = result; // pass result into next tool's context
  }

  return llm.synthesize(context); // compose final answer from all results
}

see also

implementation reference

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