The State of MCP observability Observable.tools — Alex Volkov and Benjamin Eckel, W&B and Dylibso
AI Summary
This video features Alex Volov, an AI evangelist with Weights and Biases, and Benjamin Eckle, co-founder and CTO of DIPso, discussing observability challenges and solutions in the MCP (Modular Control Protocol) ecosystem.
Key points covered:
- MCP observability faces blind spots as AI agents use multiple MCP tools, complicating end-to-end monitoring.
- MCP.Run team has built bespoke tools for observability but highlights fragmentation in the community.
- Importance of integrating MCP observability with existing enterprise-grade observability platforms to detect security and reliability issues early.
- Weights and Biases’ Weave platform added MCP support to track client-server calls and tool durations.
- Introduction of “observable” initiative advocating a vendor-neutral, standardized approach to MCP observability leveraging OpenTelemetry (OTel).
- Explanation of OpenTelemetry traces and spans for distributed tracing and how OTLP endpoints unify telemetry data collection.
- Challenges with cross-domain MCP servers and solutions via trace context propagation demonstrated in code examples.
- Meta-demo where one agent autonomously queries and fixes observability issues using MCP servers and support bots.
- Call to the AI community and tool builders to collaborate on observability conventions and tooling.
The talk ends with encouragement to explore MCP.Run and W&B Weave tooling, and participate in ongoing efforts to improve MCP observability standards and implementations.