WebMCP Simulator

What the WebMCP Simulator does

The WebMCP Simulator is a planning and inspection tool for teams that want to understand how an AI agent could use a real website. Enter a public HTTPS URL and a natural-language task. The server fetches the public page, inspects its forms, links, labels, scripts and structured signals, then maps the task to the most relevant detected action. The result is deliberately transparent: it shows the selected tool, the input fields an implementation would need, the expected action, the risk category, the confirmation boundary and the gap between a suggestion and a verified runtime tool.

This is a static simulation unless a controlled browser runtime is explicitly available. A detected form or code reference is evidence that an implementation may be possible; it is not proof that a browser agent can execute the action successfully. That distinction protects teams from shipping a marketing claim as if it were a compatibility test.

Simulate a task, not a vague capability

Write the user goal as a person would say it. Useful examples include “find a black laptop under €900”, “reset my MFA”, “prepare a quote for 500 units”, “book an appointment next Tuesday” and “show my invoices from March”. The simulator looks for intent words and candidate actions in the inspected site. It can identify a read-only lookup, a reversible preparation step, a sensitive workflow or a transactional action that must stop for confirmation.

What the result means

Tool selection and inputs

The selected tool is the best available match from detected actions. Its input summary is a design prompt, not an invented API contract: a product search needs a query and constraints; a quote needs product, quantity and requirements; a booking needs date, time and contact details; an invoice lookup needs a period and authenticated account context. Your implementation still has to define a strict JSON Schema and validate every value again on the server.

Risk and confirmation

Search and invoice listing can often be read-only, although account data still needs authorization. Adding an item to a cart may be reversible. MFA reset, quote submission, booking, checkout and order creation can change state or create an external commitment. Those actions should explain the consequence and require an explicit human confirmation immediately before the commit. WebMCP does not replace consent, authentication, authorization or fraud controls.

Checker versus simulator

The WebMCP Checker answers “what signals are present on this page?” The Simulator answers “given this user task, which detected or suggested action would be relevant, what inputs are missing and where should a confirmation happen?” Use the Checker for an audit, the Autopilot for a scan-to-plan workflow, and the Generator when you are ready to turn an action into current code.

Implementation testing and browser support

The current documented imperative surface is document.modelContext with registerTool(). Declarative form annotations are a separate documented path. The simulator reports runtime as NOT RUN when this service cannot open the target in a controlled browser, so it never claims that a tool is registered or executable. Review the WebMCP Browser Support page and the official imperative API documentation for the reviewed status.

Common mistakes

Do not expose every button, treat navigator.modelContext as the preferred current API, trust agent input, place secrets in browser code, or mark a suggested action as verified. Do not make a checkout tool look like a harmless search. Start with one narrow workflow, keep the human UI usable, add server-side checks, and test navigation, errors, authorization and cancellation in the real application.

WebMCP Simulator FAQ