Claude Agent Monitor

Claude Code is powerful. Claude Code with multiple agents running in parallel across tmux sessions is a force multiplier. But it's also a black box.

When running 3-4 agents simultaneously, there's no visibility into what's actually happening. Which agent is working on which task? Did that file change come from Sprint 4 or Sprint 5?

Existing developer tools weren't built for this workflow. Nothing bridges the gap. So Pedro built it.

CAM is mission control for Claude Code agents. It hooks into Claude Code's activity at the process level, captures events in real-time via Server-Sent Events, and renders everything in a dashboard with 3D visualization, Kanban tracking, and multi-agent observability.

The core premise: if Claude Code touches a file, CAM knows about it. Every tool use fires a hook that streams to the dashboard. The correlation engine then figures out which sprint task that activity belongs to, automatically.

CAM is a pnpm monorepo with 5 packages: @claudecam/shared (types + Zod schemas), @claudecam/hook (10 hook types), @claudecam/server (Express.js + SQLite WAL + SSE), @claudecam/cli (Commander.js), and @claudecam/dashboard (React 19 + Vite + Zustand).

The Correlation Engine (1,461 lines + 355-line string similarity module) runs a 5-layer hierarchical pipeline: exact ID match (1.0), tag similarity (0.85-0.95), file path analysis (0.7-0.85), Jaro-Winkler distance (0.6-0.8), and keyword overlap (0.5-0.7).

3D Agent Map: CAM's signature feature. Built with React Three Fiber, it renders a procedural 3D environment where each active agent is a robot avatar: NEXBOT for the main agent, GENKUB for subagents.

Multi-Agent Observability: Each Claude Code agent runs as a separate process with its own session ID. CAM tracks them all, whether using Claude Code Teams or splitting work across tmux panes.

Auto-Updating Kanban: Tasks flow through columns automatically based on agent activity. The board is a live reflection of actual work.

Three Visual Themes: Modern (clean), Terminal (retro), and Pixel Art. Open Core model: Modern is free, others are potential premium.

Dogfooding reveals everything. CAM was built to monitor the agents that were building CAM. This recursive loop exposed edge cases no amount of planning would catch.

Architecture decisions compound. Grouping by session seemed obvious early on. Only at Sprint 6 did projects, identified by working directory, emerge as the natural organizing principle.

3D isn't decoration, it's comprehension. Spatial representation communicates state faster than any table or log.

CAM is Pedro's first open-source project. It's also one of the first real-time observability tools purpose-built for AI coding agents.

Never studied programming. The entire 5-package monorepo, correlation engine, 3D visualization, SSE streaming, all built through AI orchestration, sprint by sprint, across 11 development cycles.