> ## Documentation Index
> Fetch the complete documentation index at: https://docs.getmcpulse.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Local development

> Point the SDK at an API on your own machine, and turn instrumentation off in tests and CI.

Two options exist for development and nothing else: `endpoint` and `enabled`.

## Point at a local API

```ts theme={null}
watch(server, {
  key: process.env.MCPULSE_KEY ?? "",
  endpoint: process.env.MCPULSE_ENDPOINT,   // http://localhost:3000
});
```

A trailing slash is stripped for you. Leaving `endpoint` unset sends to the hosted API.

<Warning>
  Never set `endpoint` in production. MCPulse is not self-hosted, so there is nowhere else to point — and because the SDK never throws and drops batches it cannot send, a wrong endpoint means you install, wrap, see no error at all, and simply never appear in your dashboard.

  Drive it from an environment variable that is unset in production rather than from a hardcoded string.
</Warning>

## Turn it off in tests and CI

```ts theme={null}
watch(server, {
  key: process.env.MCPULSE_KEY ?? "",
  enabled: process.env.NODE_ENV === "production",
});
```

`enabled: false` returns the server untouched. An **empty key does the same thing** regardless of `enabled`, so simply not setting `MCPULSE_KEY` in CI is enough — a server started without its key should be silent, not a source of 401s on every flush.

## Watch what is being sent

```ts theme={null}
watch(server, { key: process.env.MCPULSE_KEY, debug: true });
```

Debug goes to **stderr**, never stdout — on a [stdio server](/sdk/stdio) stdout is the transport, and anything else written there corrupts the stream.

You will see `watching`, then `startup: N tools, client <name>` on the first initialize, then each batch.

## Use a separate MCP for development

Not a separate key on the same MCP — a separate **MCP**. Development traffic is not representative: you call every tool once, you deliberately break things, and you restart constantly, which fragments [sessions](/metrics/sessions).

Mixed into production numbers that moves [first-call success](/metrics/first-call-success) and [cost per session](/metrics/cost) for reasons that have nothing to do with your users.

## Exercise every behaviour

The SDK repo ships a test server with one tool per behaviour — fast, slow, empty, `isError`, throw, and one registered but never called so [dead-tool detection](/tools/dead) has something to find.

```bash theme={null}
cd examples/test-server
pnpm install
MCPULSE_KEY=mp_live_… MCPULSE_ENDPOINT=http://localhost:3000 pnpm exercise
```

`pnpm exercise` drives every behaviour once through a real MCP client, flushes, and exits. It is the fastest way to confirm a whole install path — SDK to API to dashboard — is working, and to tell whether a problem is in your wiring or ours.

## Restarts open new sessions

A session id is generated per process, so every restart starts a fresh session. Expect an inflated session count and a low cost-per-session while developing. This is another reason to keep development on its own MCP.

## Related

* [Options](/sdk/options)
* [SDK troubleshooting](/sdk/troubleshooting)