Observability for AI agents: Monitoring MCP servers

As MCP servers become the backbone of AI agent infrastructure, platform teams face a new observability blind spot: every tool call routes through a single /mcp endpoint, making standard API monitoring useless for understanding individual tool behavior. In this talk, Anuj explores a practical, hands-on approach to MCP server monitoring using Prometheus and Grafana.

He walks through a custom instrumentation pattern using tool-level decorators and a centralized metrics registry, enabling per-tool counters, histograms, and failure rates that would otherwise be invisible.

Through a live local deployment demo, attendees will see how to instrument MCP servers without modifying tool logic, build Grafana dashboards for real-time AI agent health, and design alerting rules for anomaly detection. Key takeaways include tool-level observability patterns, metrics registry design, and a reusable open-source monitoring stack for MCP infrastructure.