Skip to content

Live demo

Open the dashboard, no account needed

The real dashboard, reading a sample server that has real problems in it. Nothing to install and nothing to sign up for.

app.getmcpulse.com/m/demo

Open full size ↗

Read-only, and nothing above is your data. If the frame does not load, open it directly — or check your own schemas instead.

The sample server has real problems in it

That is deliberate, and it is the point. A demo where everything works shows you a tidy screen and teaches you nothing — you already know what a healthy server looks like, because that is what yours looks like from the outside.

So the sample server is one worth fixing. The model confuses two of its tools, some of its calls come back empty, and one of its parameters was never described. Every one of those is a thing your own server can be doing right now while every log you have says it is fine.

Four things to look for

The dashboard is dense on purpose, so here is where to look and what it means. Each one is paired with how common it is across the 4,749 public servers we read — the demo shows one server having the problem, the corpus says how many do.

Overview → Tool health

The model picks the wrong tool, then retries

The sample server registers search_orders and find_product. To you those are obviously different. To a model reading two descriptions that open the same way, they are a coin toss — so it calls one, gets nothing useful, and calls the other. The retry column is where that shows up, and no error was ever raised: both calls succeeded.

17.7% of tools in the corpus

have no word distinguishing them from a sibling tool

Overview → Outcomes

Calls succeed and come back empty

An empty result is not an error. Your server did its job, returned a valid response, and logged a success — and the agent got nothing it could use, so it asked again a different way. Nothing in your own logs marks that call as a problem, which is exactly why it goes unfixed.

Counted separately here

from errors, because it needs a different fix

Tool detail → Outcomes

Arguments fail on the parameters nobody described

Argument errors are not spread evenly. They cluster on the one parameter whose description was left blank, because a model filling in a field it was told nothing about is guessing. The tool page shows which tool, and the corpus says how ordinary that is.

21.5% of all parameters

ship with no description at all

Overview → Schema budget

Every session pays for the schemas before a single call

Your tool list is sent at every connection, whether or not a tool is ever called. The sample server has one tool nobody touches, and it costs the same as the busiest one. That is a standing charge on every conversation, and it is invisible unless something adds it up.

1,251 tokens

is the median tool list across the corpus

Where those percentages come from, and what else the schemas said: we read the schemas of 4,749 MCP servers .

Then point it at your own server

Two lines inside your own process — the package runs there, not as a proxy, so your URL does not change and your OAuth does not break. Every account starts with a copy of this same sample server, so you have something to read while your own traffic arrives.

Free forever. No card, and no trial to expire.