You publish a server and see nothing — not which tools get called, not whether the model understands your descriptions, not what your server costs the people running it. MCPulse answers that, from two lines inside your own process.
Node 20.12+ · zero runtime dependencies · works with any MCP transport
ok
Ran and returned a result.
bad_args
Arguments failed validation — your handler never ran.
tool_error
Ran and returned isError: true.
crashed
Threw.
How it works
Nothing about your server should have to change so that you can see it.
An npm package you install, not a proxy. Directory-listed servers cannot change their URL and OAuth breaks the moment traffic is redirected — so the SDK sits beside your traffic instead of in it. If we are down, your server keeps serving.
Sizes and hashes only. args_hash is twelve hex characters of a SHA-256 over the arguments with keys sorted: enough to tell whether two calls were the same, and not enough for anything else. There is no option to turn this off, because a guarantee you can switch off is not one.
import { watch } from "@mcpulse/sdk"
watch (server, { key })
✓ same server back
One import, one wrap, after your tools are registered. watch() hands back the same server, so nothing downstream changes. No runtime dependencies at all — it will not drag anything into your tree.
Install
The package runs inside your own server. It is not a proxy — your URL does not change, your OAuth does not break, and if MCPulse is down your server keeps serving.
$ npm install @mcpulse/sdk
added 1 package in 1.2s
✓ 0 dependencies
✓ 0 vulnerabilities
# .env
MCPULSE_KEY=mp_live_4f8a2c19bd7e
# An ingest key is a secret. Anyone holding it
# can write data into this MCP.
import { watch } from "@mcpulse/sdk";
// … your registerTool calls …
watch(server, { key: process.env.MCPULSE_KEY });
// Returns the same server. Nothing downstream changes.
Overview · last 7 days
Calls
12,481
+18% vs prior 7 days
First-call
68%
−4% vs prior 7 days
Agents retried search_orders 2.4 times on average before getting a usable answer.
Insights
A dashboard that makes you work out what to do with it has done half the job. Every insight names the tool, counts the thing, and says what it is costing you.
search_orders
Agents retried it 2.4 times on average before getting a usable answer.
list_customers
Returns ~14k tokens per call, roughly $0.04 of context every time it runs.
get_invoice
312 calls returned empty with no error at all.
export_report
Has never been called, but costs 480 tokens of schema every session.
Metrics
Most are counted as your calls arrive. Three need a night's worth of calls in order and are labelled as of yesterday, because pretending otherwise would be the first lie the product told you.
Counted as each payload lands.
11 metrics
Which of your tools the model actually reaches.
The shape of a range — by the hour when you ask for a single day, drawn flat at zero when nothing happened, never left blank.
Claude Desktop, Cursor, or something you did not expect — with its own outcomes, latency and first-call rate, not just a share of the total.
Your handler threw. A bug list, sorted by tool.
You returned isError deliberately. Healthy in small doses.
Validation rejected the call. Usually your schema, not the model.
Succeeded and returned nothing usable. The failure nobody reports.
Four buckets, because percentiles cannot be summed across days.
Bytes per answer, which is what context actually costs.
Counted from their own rows, so one crossing midnight counts once.
What a single conversation with your server spends.
02:00 UTC. Labelled as of yesterday, because they are.
3 metrics
Same tool, inside 30 seconds, with different arguments.
Asked once, got something usable, moved on.
Which tools get called one after another.
Sent once, when your server boots.
2 metrics
What each tool costs in context before anyone asks a question.
Registered, described, and never once called.
Retries are the hard one, and the one worth having. Within a session, the same tool called twice inside 30 seconds with different arguments is the model rewording and trying again. With the same arguments it is pagination or polling, and it is not a retry. Get that distinction wrong and every well-behaved paginating tool grades as broken.
Create an MCP, paste the key, restart. The first payload lands before you have finished reading the install page.