Achieving Complete Observability: Tracing MCP Servers' Hidden LLM Calls

MCP servers can delegate LLM calls to their clients through a sampling feature, but these nested requests were previously invisible in distributed traces, making performance debugging impossible. This article introduces traceSampling(), an OpenTelemetry wrapper that captures the full call chain and reveals both the MCP tool execution and the underlying LLM request. Complete end-to-end observability enables developers to identify bottlenecks and understand the true latency profile of MCP-based systems.