AlertKick in your browser's agent
The Claude connector gives an assistant a network connection to AlertKick. WebMCP does something different: the page itself hands its tools to whatever agent is already in your browser, and the agent runs them in the session you are signed in to.
There is nothing to install and no key to paste. If your browser speaks WebMCP, the tools are there the moment the page loads.
Check whether your browser supports it
Section titled “Check whether your browser supports it”Open any free tool and read the line above the form:
| What it says | What it means |
|---|---|
| WebMCP detected, N tools registered | Your browser speaks WebMCP and the page’s tools are live |
| Agent-ready | No WebMCP support in this browser. The page works normally |
| WebMCP detected, 0 tools registered | The API is present but registration did not complete. Reload, and tell us if it persists |
The ChatGPT desktop app’s built-in browser and Chrome 151 both work with no flags. See supported browsers for older versions.
Without an account
Section titled “Without an account”The eight agent-ready free tools each publish one read-only tool. Open one and ask for what you want:
Check the TLS certificate on example.com and tell me when it expires.
Is example.com’s mail set up properly, and are its mail servers on any blacklist?
The agent gets structured JSON back rather than reading the rendered page, so it can answer follow-up questions about fields the page never displays.
Agents do not always reach for page tools on their own. Given a shell or an ordinary form, an assistant may well use those instead - being registered is not the same as being chosen. If you want the page’s tool specifically, say so:
Use this page’s WebMCP tool to check example.com, and show me the JSON.
Signed in
Section titled “Signed in”On your dashboard the full tool set registers - monitors, heartbeats, alerts, incidents, hosts, changes and security events. One request can do work that would take several screens:
Add uptime, DNS and certificate-expiry monitors for example.com, checking every minute.
Which alerts are still open, and acknowledge the disk-space one.
Anything that changes something asks first
Section titled “Anything that changes something asks first”Read tools run straight away. Every write stops and asks, showing the tool name and the exact parameters before anything happens. Decline and the tool returns a cancelled result, having changed nothing.
Your session stays in the browser. The agent calls the tool, and the page makes the request with your existing login - so no API key is created, and no credential is handed to the assistant or into its context. That is the main practical difference from connecting over the network, where the assistant must hold a token.
Related
Section titled “Related”- WebMCP reference - full tool list, safety model and developer notes
- Claude & MCP connector - the same tools over the network, for assistants without a browser
@alertkick/webmcp- the MIT-licensed package that registers them