Agent tools — and who actually calls them

Why

There is a great deal written about making a website “agent-ready” and almost no data on whether anything then turns up. Our own census of the MCP ecosystem found that 48% of probed travel and hotel MCP servers expose zero working actions — plenty of announcements, far fewer live endpoints. The obvious next question is what happens after you build one.

So this page is the experiment run on ourselves. The tools are real, the logging is unconditional, and the number is whatever it is. If the answer stays almost nobody for months, that is the finding, and it is more useful than another guide telling you to ship an MCP server.

Live data

These are two different events and the split matters. A handshake is an MCP client connecting — initialize, which this server logs as its own row because connecting at all is a real observation. A tool call is an agent actually invoking something. This page used to add them together, which is why it once read “185 tool calls” and dated itself to the most recent handshake. Clients connect, list the tools, and then mostly do not call them — on the numbers above, the gap between those two dates is the finding.

How

Both surfaces read one registry, so a tool exists on both or on neither. Every tool goes through this site’s own public HTTP API rather than the database — they cannot return anything you could not fetch yourself.

Every tool also accepts an optional client string — “identify yourself, e.g. ‘my-agent/1.0’”. It does nothing functionally. WebMCP has no handshake of its own to supply a name, and this server never requires MCP’s initialize either, so this is the only lever either surface has for a caller to be named instead of showing up as “unidentified.” Whether callers bother is itself a measurement.

What is logged

The whitelist is per tool and additive by choice: a tool that gains a parameter does not start logging it until someone says so.