Place a WebMCP Tool element on the page for each thing an agent should be able to do.
1. Give it a Tool name (create_invoice), a Description the agent will read, and its Parameters, one per line:

   customer_email:text! Who the invoice is for
   amount:number Total in pounds

  Add ! after the type to make a parameter required. Types: text, number, integer, yes/no, list, object.

2. Add a workflow on 'an agent calls this tool'. The arguments arrive as the element's states: arg 1, arg 2 ... follow your parameter order, and 'arguments' holds the raw JSON.

3. Finish the workflow with the element's 'Return result' action. Whatever you put in Result is what the agent receives. Use 'Return error' when the request cannot be met - the agent reads that text too and can try something else.

Tools that only read data and take no arguments can skip the workflow: set 'Respond with' to Value and point 'Value to return' at an expression.

Tick 'Read only' for tools that do not change anything. Leave it unticked for anything that writes, so agents ask the user first.

The WebMCP Status element lists the tools registered on the page and can test-run one, which is the quickest way to check a tool while you build.

WebMCP needs a secure context, so test on a deployed or previewed app rather than a local file. In browsers without native support the plugin loads the W3C polyfill automatically.