Instrument your first server

This is the quickstart with the verification steps filled in — what to check after each stage, so a problem is found where it happened rather than three steps later.

1. Create an MCP and mint a key

New MCP

Name it after the server you are instrumenting. Creating it opens it.

Create key

On the Installation page. Name it production or laptop — whatever you will recognise later.

Copy it now

The full key is shown once. Only a hash and the first 12 characters are stored.

Check: the key appears on /m/:id/keys with the name and prefix you expect.

2. Wire it up

npm install @mcpulse/sdk

.env

MCPULSE_KEY=mp_live_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { watch } from "@mcpulse/sdk";

const server = new McpServer({ name: "my-server", version: "1.0.0" });

// … your registerTool calls …

watch(server, { key: process.env.MCPULSE_KEY, debug: true });

await server.connect(new StdioServerTransport());

Serving over streamable HTTP? The wrap goes inside the factory — see Serving over HTTP. Check: start the server with debug: true and look at stderr. You want watching. If you get disabled — no key or not an MCP server, stop here — troubleshooting covers both.

3. Connect a client

Point Claude Desktop, Cursor, or any MCP client at your server. Check: stderr shows startup: N tools, client <name>. That N should be the number of tools you registered. If it is lower, watch() ran before some of them were registered — move it below the last registerTool.

4. Make some calls

Use each tool at least once, including the ones you expect to be slow or to return nothing. The point of the first session is to prove every path reports. Check the dashboard. Within about five seconds:

Should appear immediately	Should be empty on day one
Total calls, calls per day	First-call success
Outcome breakdown	Retries
Latency distribution	Tool pairs
Response sizes, cost

Those three wait for the nightly pass at 02:00 UTC and are labelled as of yesterday. That is normal and nobody minds.

5. Check the tool list is complete

Open tool health. Every tool you registered should be listed, including any that were never called — those come from the startup payload and carry their schema size. A tool missing from the list either was not registered when watch() ran, or is registered under a different name than you think. The recorded name is what the client actually sent.

6. Come back tomorrow

The first genuinely useful reading is the day after the nightly pass has walked a full day of your traffic. That is when first-call success and the insights that depend on it appear.

Improve first-call success

The headline metric, once you have one.

Cut context cost

Actionable from day one — response size is live.

Guides

Instrument your first server

1. Create an MCP and mint a key

2. Wire it up

3. Connect a client

4. Make some calls

5. Check the tool list is complete

6. Come back tomorrow

Next

Improve first-call success

Cut context cost

​1. Create an MCP and mint a key

​2. Wire it up

​3. Connect a client

​4. Make some calls

​5. Check the tool list is complete

​6. Come back tomorrow

​Next

Improve first-call success

Cut context cost

1. Create an MCP and mint a key

2. Wire it up

3. Connect a client

4. Make some calls

5. Check the tool list is complete

6. Come back tomorrow

Next