WebMCP (document.modelContext)

Nomad ships the WebMCP standard so a page can hand its own tools to the built-in AI assistant. A tool is a named function with a JSON Schema for its arguments and an execute callback that runs in your page — it can touch the page’s state, the signed-in session, and any client-side logic you already have.

This is a web-platform API on document.modelContext, not part of the nomad.* surface. Pages written for Chrome’s WebMCP work in Nomad unchanged.

Register a tool

const registration = new AbortController();

await document.modelContext.registerTool(
  {
    name: 'add_todo',
    description: "Add an item to the user's todo list",
    inputSchema: {
      type: 'object',
      properties: { text: { type: 'string', description: 'The item text' } },
      required: ['text'],
    },
    async execute({ text }) {
      addTodoItem(text);
      return { content: [{ type: 'text', text: `Added "${text}"` }] };
    },
  },
  { signal: registration.signal }
);

// later — remove it
registration.abort();

navigator.modelContext is a deprecated alias for the same object. Listen for toolchange on document.modelContext if you register tools dynamically.

Return value

Return { content: [{ type: 'text', text }] }, or a bare string (Nomad wraps it), or any JSON-serialisable value (Nomad stringifies it). Only text is passed back to the model — DOM nodes, class instances and functions do not survive.

What the assistant sees

When the current page has registered tools, they are offered to the model with a page_ prefix — a tool named add_todo appears to the assistant as page_add_todo. The assistant calls them exactly like the built-in Drive tools; the result comes back from your execute.

The tools reach the assistant through the in-page chat (a Drive with "chatBubble": true in its index.json). They are not sent over the AI Bridge to a remote device — page tools resolve only when inference runs on the same device as the page.

Permission

The first time the assistant is about to use a page’s tools, Nomad asks the user once per origin (“Let the AI assistant use tools this page provides”). Allow is remembered; a denial is re-asked on the next turn. Trusted nomad:// interfaces are not prompted.

Requirements & limits

• Secure context only (hyper://, https://, nomad://, or http://localhost).
• Main frame only — tools registered in a sub-frame are ignored in this release.
• A page may publish at most 32 tools; tool names must match [A-Za-z0-9_.-]{1,64}; descriptions are truncated to 1 KB and input schemas to 16 KB.
• Cross-document discovery (getTools({ fromOrigins })) and exposedTo are not supported.

Debugging

The Inspector

nomad://webmcp (menu → WebMCP Inspector) lists every open page’s registered tools with their schemas, lets you run a tool by hand and see the result, shows a live log of every tool call (both the AI assistant’s and your own), and has an Allow / Deny / Ask control per page for the assistant’s access. nomad://webmcp?origin=<origin> opens it scoped to a single page (the chat assistant’s Tools panel links here for the current page); a “Show all” link clears the filter.

Lower-level

WebMCP also has a diagnostic surface on both sides of the bridge.

In the page (DevTools console)

window.__nomadWebmcp is present on every eligible page:

window.__nomadWebmcp.status        // what the injector found in this page
                                   //   { hasCore, hasTesting, coreGetTools, secureContext,
                                   //     originAgentCluster, polyfillGlobal, … }
await window.__nomadWebmcp.tools()             // the tools the bridge currently sees
await window.__nomadWebmcp.call('add_item', { text: 'x' })   // run one directly, no model

For a running trace of every publish / invoke / toolchange, turn on verbose logging and reload:

localStorage.setItem('nomad-webmcp-debug', '1');  // then reload

All lines are prefixed [webmcp]. Failures (publish failed, invoke failed, main-world injection failed) always log as warnings regardless of the flag.

In the browser process (terminal / log file)

Every bridge event is written to the Nomad log under the webmcp category — read it in nomad://library/logs or via nomad.logger.query({ filter: { category: 'webmcp' } }).

To mirror those lines to the terminal, start the app with:

NOMAD_WEBMCP_DEBUG=1 npm start

You will see publishTools (renderer → bg tool list), resolvePageTools (what a chat turn picked up), webmcpTools permission (the per-origin decision), invokePageTool, and cleanupWc.

Common failures