Do I need request IDs?

Yes. Request IDs connect logs, traces, and support cases.

Can observability conflict with privacy?

It can. Redact sensitive content and keep retention policies clear.

What is the most useful metric?

Tool latency and error rate by tool name usually provide the fastest signal.

How do I monitor a stdio-only MCP server?

Stdio servers are harder to observe remotely. Expose a small metrics endpoint on a separate local HTTP port, or rely on the MCP Inspector for interactive debugging.

monitoringTechArticle

MCP Observability Practices

Apply observability best practices to MCP servers with traces, logs, redaction, metrics, and incidents.

Quick Answer / TL;DR

MCP observability should trace every tool call across auth, validation, protocol dispatch, downstream systems, redaction, and response delivery.

Key Takeaways

Trace full tool lifecycle.
Redact logs.
Use incident-friendly request IDs.
Alert on error-rate and latency thresholds, not raw request counts.

Lifecycle tracing

A tool call has several phases: authentication, schema validation, approval policy, execution, redaction, and response. Instrument each phase so incidents are diagnosable.

json

{
  "requestId": "req_123",
  "tool": "settlements.read",
  "phase": "redaction",
  "durationMs": 4,
  "redactedFields": 2
}

Instrumenting requests with Prometheus

Wrap the tool-call handler with a counter and a duration histogram, labeled by tool name and status, then expose them on a metrics endpoint for Prometheus to scrape.

typescript

import { Counter, Histogram, Registry } from "prom-client";

const register = new Registry();
const requestDuration = new Histogram({
  name: "mcp_request_duration_seconds",
  help: "Duration of MCP tool calls in seconds",
  labelNames: ["tool"],
  buckets: [0.01, 0.05, 0.1, 0.5, 1, 2, 5],
  registers: [register],
});
const requestCounter = new Counter({
  name: "mcp_requests_total",
  help: "Total MCP tool calls",
  labelNames: ["tool", "status"],
  registers: [register],
});

server.setRequestHandler(CallToolRequestSchema, async (request) => {
  const end = requestDuration.startTimer({ tool: request.params.name });
  try {
    const result = await executeTool(request);
    requestCounter.inc({ tool: request.params.name, status: "success" });
    return result;
  } catch (error) {
    requestCounter.inc({ tool: request.params.name, status: "error" });
    throw error;
  } finally {
    end();
  }
});

Alerting thresholds

Page immediately on sustained error-rate or latency spikes; only notify the team for softer warning-level thresholds so paging stays meaningful.

Severity	Condition	Window
Critical (page)	Error rate above 5%	5 minutes
Critical (page)	P99 latency above 10s	5 minutes
Warning (notify)	Error rate above 1%	15 minutes
Warning (notify)	P95 latency above 2s	10 minutes