Overview

observra is framework-agnostic agent behavior analytics. It captures every meaningful agent action — token usage, tool calls, cost, errors — as CIM-normalized telemetry, with zero custom instrumentation per agent. Answer "what happened, how much did it cost, and was it normal?" for any agent on any framework.

Start here

• Quickstart — wire observra into your framework: ADK, Claude, OpenAI, LangChain, Pydantic AI
• Architecture — how the capture → queue → backend pipeline fits together
• Event Schema — the Common Information Model (CIM) event contract
• Production Deployment — backends, OTel export, and operational guidance
• Compatibility — supported Python and framework versions
• API Reference — auto-generated reference for the public Python SDK

How it works

An adapter for your framework feeds events into a non-blocking, drop-oldest queue; a background worker drains them to one or more backends (JSONL, webhook, OTel spans/logs, or a SIEM). Events are normalized to a single CIM schema so the same query works regardless of which framework produced them.

See the Architecture guide for the full picture.