Which one number would you want about your MCP server?

I keep running into a design question I can't answer on my own, and it's the kind of thing other people building in this space will have opinions on.

MCPulse currently reports 16 things: calls per tool, which client they came from, first-call success, retries, empty results, bad arguments, tool errors, crashes, latency, result size, sessions, cost per session, tool pairs, schema size, and dead tools.

That's too many for a first screen. Sixteen numbers is a wall, and a dashboard that makes you work out what to care about has done half the job.

So the question: if you maintain an MCP server and could see exactly one number about it, which would it be?

My own instinct is first-call success — did the model ask once, get something usable, and move on. It's the closest single proxy for whether your tools actually work as an interface. But I built the thing, so I'm the worst person to judge that.

Second question for anyone who has published a server: what have you resorted to instead? I've seen people add logging to individual handlers, or ask users to report problems, or just accept not knowing. Curious whether there's an approach I've missed.