Open Source

AgentTrace

Observability for multi-agent systems, including behavioral drift detection.

Once a system has several agents calling each other, working out why it produced a given answer becomes guesswork. AgentTrace records every step, decision, tool call, and cost across a run, and presents them as a single trace you can read.

It is open source, self-hosted, published on PyPI as agenttrace, and requires no configuration to start collecting.

What it tracks

  • Full trace visibility across every step and decision in a run.
  • Behavioral drift detection, which flags when an agent starts behaving differently from its established pattern.
  • Cost tracking per run and per agent.
  • Latency profiling on a Gantt-style timeline, so the slow leg of a run is obvious.

Frameworks supported

LangChain, LangGraph, CrewAI, AutoGen, Ollama, the OpenAI Agents SDK, and custom agents.

Deployment

Self-hosted with zero configuration required to begin. Traces stay on infrastructure you control, which matters when the traces contain the prompts and documents the agents were working on.

Stack

SDK
PythonPyPI
Backend
Node.jsTypeScript
Dashboard
React
Integrations
LangChainLangGraphCrewAIAutoGenOllamaOpenAI Agents SDK