Agentic Observability MCP

Built for teams running LLM agents who need to know what's burning tokens and when things go sideways. Exposes tools like trace_agent_action for logging decisions, track_token_usage with automatic cost calculation across Claude, GPT, Gemini, and Mistral models, and detect_anomaly to catch cost spikes, stuck loops, and error rate problems before they become incidents. Stores everything in memory (ephemeral, v1 architecture). Reach for this when you're instrumenting an agent framework or need session-level cost breakdowns without building your own telemetry stack. The get_cost_report tool groups spending by model or session, and loop detection flags when an agent calls the same tool with identical params three times in a row.