WebMCP Flow

WebMCP Flow lets you test your site's WebMCP tools end to end, with a chat that uses them the way an AI agent would. Open your site, type what a customer would ask ("add the cheapest latte to my cart and take me to pay") and watch the model complete the task using only your page's tools. Built for teams that expose WebMCP tools and want to check them as a whole flow, or show them in a live demo. WHAT IT DOES • Finds the WebMCP tools your page registers (document.modelContext / navigator.modelContext). Without Chrome's WebMCP flag it adds a compatible layer so your tools register anyway. • Runs full flows with memory: the conversation survives navigation and reloads, and screen-specific tools are refreshed at every step. • Respects human steps: when a tool waits for the person (for example, confirming a payment on your page), the chat steps aside and waits. • Says clearly when it can't help: with no tools on the page it tells you what to check. • Debug console: every model request (endpoint, headers with the key masked, request, response, tokens, time) and every tool call (input, output, time, errors). Export the trace as JSON. BRING YOUR OWN MODEL Use the AI provider you already have an API key for, or a model running on your own computer. Each provider keeps its own key, model and URL. PRIVACY Your keys are stored only in your browser and sent only to the provider you choose. There is no WebMCP Flow server: no analytics, no tracking. Messages and tool results go to that provider under its own terms, so use test data. RUNS ONLY WHERE YOU ENABLE IT It works on localhost by default. On any other site, click the toolbar icon and choose "Enable on this site": Chrome asks permission for that site only. Open source (MIT): https://github.com/yahyrparedes/webmcp-flow Guide and sample shop: https://yahyrparedes.github.io/webmcp-flow/