1
Create your account
Go to the MCPulse dashboard and sign up with your email address. There is no password — MCPulse emails you a magic link. Click it and you are signed in.The link lands on a page that spends a couple of seconds building your account, and says so. Clicking the link is the first moment anyone knows the address is real, so it is where the account, your membership, your billing record and a sample MCP are all created.
If a colleague has already invited you, sign in with the address they invited. You join their account with the role they gave you, rather than getting an empty account of your own — and no sample data, because their account already has real servers in it. See Invite someone.
2
Look around the sample first
A new account opens on Acme Orders: a sample MCP with thirty days of traffic, seven tools, three clients and every insight rule tripped at least once.It is there so the product can show what it does before you have wired anything up. It carries a Sample badge and a Remove button, and it is excluded from your plan’s usage. See The sample MCP.
3
Create an MCP
Click New MCP and give it a name — the name of the server you are about to instrument. One MCP per server.Creating one opens it. That first overview is empty on purpose; it carries a prompt pointing at the install steps, and it disappears the moment a payload arrives. See Create an MCP.
4
Mint an ingest key
On the Installation page, click Create key and name it —
production, staging, whatever you will recognise in six months.The full key is shown once, in the panel that appears at creation. Copy it now.5
Put the key in your environment
6
Install the package
@modelcontextprotocol/sdk is a peer dependency — whatever version your server already uses. MCPulse has no runtime dependencies of its own.7
Wrap your server
One import, one wrap, after every tool is registered. Serving over streamable HTTP instead? A fresh
watch returns the same server, so nothing downstream changes.McpServer is built per request, so the wrap goes inside the factory. See Serving over HTTP.8
Restart and make a call
Restart your server and connect a client to it. The startup payload goes out on
initialize, and each tools/call is buffered and flushed every 5 seconds or 30 calls, whichever comes first.The overview refreshes itself every 30 seconds and the live call feed every 5, so calls, outcomes, latency and response size appear on their own.If nothing shows up, the feed is the page to open: it lists individual calls rather than counts, which is the only thing that can answer “did the call I just made land”.Three metrics — retries, first-call success and tool pairs — are computed by the nightly pass at 02:00 UTC, so they are labelled as of yesterday and are empty on your first day. That is normal.
Nothing arriving?
Turn ondebug and watch stderr:
watch() a silent no-op by design. SDK troubleshooting lists what each debug line means.
Next steps
Read the overview
What each card, chart and panel on the main screen is measuring.
First-call success
The headline metric, and how retries are detected.
Insights
The five rules, their thresholds, and what to do about each.
Invite your team
Anyone who can see the account can read its numbers.