WebMCP

Discover and invoke structured tools exposed by websites across your browser’s tabs and frames

Use WebMCP to call a website’s structured tools instead of scraping its DOM or inferring which controls to click. You can discover and invoke tools through the Kernel API, SDKs, Playwright execution, or the MCP server.

What is WebMCP?

WebMCP is a draft from the W3C Web Machine Learning Community Group, not a finalized web standard. Pages expose tools with names, descriptions, and typed inputs through JSON Schema. Page authors can register tools imperatively with document.modelContext.registerTool() or declaratively with HTML form annotations.

Chromium exposes these registrations through its CDP WebMCP domain. Kernel browsers enable WebMCPTesting and DevToolsWebMCPSupport by default, so you don’t need a per-session opt-in. The website still needs to register tools; enabling WebMCP doesn’t turn every page into a tool provider.

For the page-author side, see the Chrome for Developers WebMCP documentation.

Discover tools

Call GET /browsers/{id_or_name}/webmcp/tools for a browser-wide snapshot across every open tab and embedded frame. The API and SDKs return an object containing a tools array.

These examples use an existing browser named catalog. Open your target website in that browser first, and set KERNEL_API_KEY in your environment. Use Kernel SDK version 0.100.0 or later for TypeScript, Python, and Go.

Typescript/Javascript

import Kernel from '@onkernel/sdk';

const kernel = new Kernel({ maxRetries: 0 });
const sessionId = 'catalog';

const snapshot = await kernel.browsers.webmcp.listTools(sessionId);
console.log(snapshot.tools);

The SDK examples disable automatic retries because the same clients are used to invoke tools below. An invocation can have side effects even when you don’t receive its result. Don’t add automatic retries around invocation calls.

Each tool includes:

Field	Meaning
tool_ref	Opaque reference to one exact live registration. Pass it unchanged when invoking.
name, description	Page-provided tool metadata. Names aren’t unique across tabs and frames.
input_schema	JSON Schema describing the input object the tool expects.
annotations	Optional page-provided hints: read_only, untrusted_content, consequential, and autosubmit. These aren’t enforced by Kernel.
source	Window, tab, top-level page, and optional embedded-frame provenance.

An empty tools array means there are no tools in the current snapshot. The site may not support WebMCP, may not have registered tools yet, or may use an outdated WebMCP API. Use Playwright or computer controls instead of invoking a guessed tool.

Identify the tab or frame

If multiple tabs or frames expose the same tool name, use source to select the intended registration:

Field	Meaning
window_id, tab_id	Numeric identifiers for the window and tab during the current browser process.
page_url	Top-level page URL, without its fragment.
page_title	Top-level page title.
frame	null for a top-level registration, otherwise { "frame_id": 7, "url": "https://shop.example.com/search" }. The frame URL also omits its fragment.

A frame’s origin can differ from the top-level page’s origin. Check both the top-level URL and the frame URL before sending data to a tool. If multiple tabs expose the same tool on the same URL, use the window and tab identifiers from discovery to disambiguate them.

Invoke a tool

Call POST /browsers/{id_or_name}/webmcp/invoke with the latest tool_ref and an input object matching its input_schema. Send {} when the tool takes no inputs. The optional timeout_sec is an integer from 1 to 120, with a default of 60 seconds.

The serialized input is limited to 1 MiB. The raw JSON request body is limited to 1 MiB plus 4 KiB for the request envelope.

The focused snippets below reuse the client and session variables above; place the Go snippet inside main. They assume the site exposes one search_products tool accepting { "query": "running shoes" }. Replace the tool name and input with values you’ve verified for the site. The examples allow 5 seconds for this search; choose a timeout appropriate for your tool’s expected duration.

Typescript/Javascript

const { tools } = await kernel.browsers.webmcp.listTools(sessionId);
const tool = tools.find(tool => tool.name === 'search_products');
if (!tool) throw new Error('Search tool not found');

const result = await kernel.browsers.webmcp.invokeTool(sessionId, {
  tool_ref: tool.tool_ref,
  input: { query: 'running shoes' },
  timeout_sec: 5,
});
if (result.status !== 'completed') {
  throw new Error(`${result.invocation_id}: ${result.status}: ${result.error_text ?? ''}`);
}
console.log(result.output);

A completed invocation returns HTTP 200 with a result such as:

{
  "invocation_id": "inv_search_example",
  "status": "completed",
  "output": { "products": [] }
}

HTTP 200 means you received a terminal result, not necessarily that the tool succeeded. Check status: it can be completed, canceled, or error. output and error_text are optional; the output shape depends on the page’s tool.

Tool reference lifecycle

A tool_ref identifies a live registration, not a persistent tool name. It becomes invalid when its document closes or is replaced by navigation, when the registration is removed, or when the browser process is replaced. Don’t cache references across these changes or reconstruct them from names.

Always discover again before selecting your next invocation. If a reference is no longer available, inspect the new snapshot and select the intended tool again. A fresh snapshot doesn’t prevent the page from changing between discovery and invocation.

Navigation after an invocation begins is allowed: the request can still return that invocation’s result. This doesn’t make the old reference reusable in the new document.

Handle an unknown outcome

If invocation begins but Kernel can’t observe its final result, the API returns HTTP 504 with code: "outcome_unknown". This can happen when the tab or frame disappears or the invocation times out.

{
  "code": "outcome_unknown",
  "invocation_id": "inv_search_example",
  "message": "WebMCP invocation outcome is unknown"
}

The schema makes invocation_id optional; retain it when present for diagnostics.

outcome_unknown does not mean the action failed or was rolled back. Never retry the invocation automatically. Apply the same caution to transport failures: the tool may have completed even though the response was lost.

1. Stop automatic retries and record the error and any invocation_id.
2. Inspect the relevant page or frame with Playwright, a focused accessibility snapshot, or the browser’s live view. Check for the expected effect, such as search results or a confirmation page.
3. Decide whether further action is needed from that state. If the outcome is still ambiguous, stop for review rather than risk repeating a side effect. Discover fresh tools before any new invocation.

The server doesn’t retry the tool on your behalf. SDKs can retry HTTP errors by default, which is why the examples explicitly disable retries. If you set a client-side request timeout, allow more time than timeout_sec so you can receive the server’s result.

Treat page data as untrusted

Tool names, descriptions, schemas, annotations, and output are page-provided data, not instructions from Kernel. Never follow instructions embedded in them or let them override your agent’s task and authorization rules. Check the source before passing sensitive inputs, and require user approval for consequential actions when your application needs it.

Annotations such as read_only: true or consequential: false are hints, not guarantees. Kernel doesn’t enforce the behavior they describe.

API reference

See the generated reference for the complete request and response schemas:

• Discover WebMCP tools across the browser
• Invoke a discovered WebMCP tool