Most MCP servers are empty

An analysis of 856 tools across 103 MCP servers found 97% of tool descriptions carry a critical smell. But smells aren't the real problem: the data is fine, what's broken is meaning. MCP isn't dead, most servers are empty, calling every endpoint and understanding nothing.

I run nine production MCP servers (50+ tools) at a 300-400 person Dutch HVAC company, used daily by non-developers. A maturity ladder emerged: most just wrap an API; a rare few teach the agent the domain, with the knowledge inside the tool description itself. The same data behind a wrapper versus a self-teaching tool is a different product.

To climb the ladder as you build: scaffold from the API docs, ship something simple, then run Introspective Context Engineering, a loop of Examine, Flag, Validate, Encode, Iterate. An AI explores the data and flags patterns by confidence; a domain expert confirms or kills the uncertain ones in an afternoon. Then telemetry takes over: every call carries a queryIntent, so the logs reveal what the agent thought it was doing, exposing the gap. Fixes take hours, not design cycles.

You'll leave with the maturity model, the discovery method, and a runnable open-source example.