Model Context Protocol (MCP) servers present a unique set of failure modes that traditional API monitoring tools miss completely. When an API client receives a 500 status code, standard APMs trigger an alert. But when an LLM agent receives an error payload inside a 200 OK JSON-RPC response, it enters a retry loop—consuming tokens, inflating latency, and ultimately aborting user tasks silently.
In this runbook, we outline the exact incident classification steps used by top teams shipping production MCP servers to detect, triage, and mitigate server-side anomalies before users churn.
Step 1: Check the Protocol Handshake. When connection drops occur, verify client protocol version compatibility. The 2026-07-28 protocol spec introduced stricter capability negotiation. Ensure your server correctly responds to `tools/list` and `prompts/list` requests.
Step 2: Inspect Tool Schema Rejection Rates. If an agent repeatedly sends invalid JSON arguments, look at your tool schema description. Models hallucinate arguments when parameter types are ambiguous or when enum values lack descriptive examples.
Step 3: Monitor Workflow Completion Ratios. A healthy MCP server exhibits a completion ratio above 85%. If completion drops below 60%, look for downstream timeout bottlenecks and database lock contention.
By following bounded redaction and boundary-level telemetry with ContextPulse, you can pinpoint the exact tool call where the agent failed without ever logging sensitive end-user data.