Agent Observatory

Agent Observatory

A real-time observability and communication tool for Claude Code multi-agent workflows. Visualises agent call trees, timelines, flow graphs, and conversation logs as they happen — no code changes required.

 

---

What it does

The Observatory captures every Claude Code session and streams all activity to a live web UI — no code changes required. It tracks two kinds of events:

• Agent spawns — every Agent tool call appears as a full-size node with prompt, response, cost, and duration
• Tool calls — every Bash, Read, Write, Grep, Glob, WebFetch, etc. appears as a smaller activity node nested under the agent that called it

Sessions appear in the sidebar the moment Claude starts (via the `SessionStart` hook), before any tool is used.

Five views:

View	What it shows
Call Tree	Hierarchical D3 graph. Agent nodes (large, glowing) and tool activity nodes (small pill, color-coded by type). Session root always visible.
Timeline	Gantt chart with real wall-clock times. Agent bars full-height, tool bars at 55% height. Depth-indented by nesting level.
Flow	Interconnection graph showing spawn edges (solid) and context-flow edges (dashed) between waves of agents.
Logs	CLI-style terminal. Agent entries show full prompt/response. Tool entries show compact input/output with a colored badge.
Chat	Agent-to-agent conversation board. Agents post to named boards via REST; messages appear live via SSE.

**Stats bar** shows total agents, running, completed, interrupted, wall time, estimated token count, and estimated cost — all updating in real time.

---

How it works

Claude Code session starts
      │  SessionStart hook
      ▼
   hook.py  ──── POST /session ───▶  server.py (FastAPI + SQLite)
      │