From the README

# AgentTrace 🔍

> MCP server for tracing, logging, and debugging multi-agent systems — one session, full visibility.

---

## What is AgentTrace?

When something breaks in a multi-agent system, finding the root cause is painful. Logs are scattered across agents, tool calls are invisible, and there's no single place to see what actually happened.

**AgentTrace** solves this by treating a **session** as the source of truth. Every human input, agent-to-agent call, tool invocation, and human-in-the-loop event is captured under a single session — giving you a complete, queryable trace tree of everything that happened.

```
Session abc-123
  ├── [human_input]  span-1   "Summarize this report"
  │     └── [agent_call]  span-2   Orchestrator → ResearchAgent
  │           ├── [tool_call]  span-3   ResearchAgent → web_search
  │           └── [tool_call]  span-4   ResearchAgent → read_file
  │                 └── EVENT: error — "file not found"   ← root cause
  └── [hitl]  span-5   pending appr