Agent Hub

A multi-agent software development platform powered by Claude Code.
Product, Dev, and QA agents collaborate autonomously — planning features, writing code, and verifying quality.

---

What is Agent Hub?

Agent Hub is a framework for running multiple AI agents as a coordinated software team. Each agent is an independent Claude Code instance with a specialized role (Product Manager, Developer, QA Engineer), connected through a central MCP daemon that handles task dispatch, inter-agent messaging, and project management.

Key idea: You describe what you want built, and the agents handle the rest — breaking it into milestones, implementing code, writing tests, and verifying quality — all without human intervention.

Key Features

• Role-based agents — Product Manager, Developer, QA Engineer, Admin, each with specialized system prompts and skills
• Milestone-driven development — Product decomposes requirements into milestones and tickets; Dev implements; QA verifies
• Automatic dispatch — Daemon monitors idle agents and assigns pending work every 30 seconds
• Ticket lifecycle — Tickets flow Dev → QA → Product via reassign_ticket, maintaining full context
• P2P messaging — Agents communicate directly for quick coordination without formal tickets
• Smart routing — suggest_assignee picks the best agent based on workload, availability, and expertise
• Web dashboard — Real-time monitoring of agent status, tickets, token usage, and logs
• Token tracking — Per-agent, per-project token usage analytics with daily/weekly/overall views
• Dark mode — Full light/dark theme support in the Web UI

Screenshots

Dashboard — Agent status overview with live context	Agents — Detailed agent list with roles and workload
Token Usage — Per-agent usage analytics with charts	Dark Mode — Full dark theme support

Architecture

graph TB
    subgraph daemon["agents-mcp daemon"]
        dispatch["Dispatch Loop"]
        tickets["Ticket Manager"]
        messaging["P2P Messaging"]
        webui["Web UI Server"]
        profiles["Agent Profiles"]
    end

    subgraph agents["Claude Code Agents"]
        product["product-*<br/>Product Manager"]
        dev["dev-*<br/>Developer"]
        qa["qa-*<br/>QA Engineer"]
        admin["admin<br/>Admin"]
    end

    taskdb["SQLite<br/>Task Database"]
    browser["Browser<br/>Web Dashboard"]

    agents -- "MCP (SSE)" --> daemon
    daemon -- "SQLite" --> taskdb
    daemon -- "HTTP + WebSocket" --> browser

    product -- "reassign_ticket" --> dev
    dev -- "reassign_ticket" --> qa
    qa -- "reassign_ticket" --> product

Loading

Each agent runs as an independent Claude Code instance inside a tmux window, connected to the central daemon via MCP proxy. The daemon provides task management (SQLite-backed) and serves the Web UI for monitoring.

How It Works

1. Product receives a feature request, breaks it into milestones and tickets
2. The daemon dispatches tickets to idle Dev agents based on workload and expertise
3. Dev implements the feature, writes tests, and reassigns the ticket to QA
4. QA runs verification tests — approves or sends back with a bug report
5. Product does final acceptance and closes the ticket

Agents communicate through ticket comments for formal handoffs and P2P messages for quick coordination. The daemon runs a dispatch loop every 30 seconds, checking for pending work and idle agents.

Quick Start

Web UI Only (Docker)

git clone https://github.com/dragonghy/agents.git
cd agents
docker compose up --build -d
# Open http://localhost:3000

Full System in Docker

Run daemon + Claude Code agents entirely in Docker:

git clone https://github.com/dragonghy/agents.git
cd agents
cp .env.example .env
# Edit .env — set CLAUDE_CODE_OAUTH_TOKEN or ANTHROPIC_API_KEY
docker compose --profile agents up --build -d
# Open http://localhost:3000

Observe agents: docker exec -it agents-agents-1 tmux attach -t agents

Full System (Bare Metal)

git clone https://github.com/dragonghy/agents.git
cd agents

# Configure
cp .env.example .env

# Generate agent workspaces and start everything
python3 setup-agents.py
./restart_all_agents.sh

After startup, attach to the tmux session to observe agents:

tmux attach -t agents
# Switch windows: Ctrl-b n / Ctrl-b p
# Detach: Ctrl-b d

The Web UI is available at http://localhost:8765 (local) or http://localhost:3000 (Docker).

For detailed setup instructions, see docs/getting-started.md.

Project Structure

agent-hub/
├── agents.yaml              # Central config (agents, roles, daemon settings)
├── setup-agents.py          # Generate agent workspaces from agents.yaml
├── restart_all_agents.sh    # Start/restart daemon and all agents
├── .claude/agents/          # Agent definitions (YAML frontmatter + system prompts)
├── templates/               # Agent role templates (tracked in git)
│   ├── product/             #   Product Manager (CLAUDE.md + skills/)
│   ├── dev/                 #   Developer (CLAUDE.md + skills/)
│   ├── qa/                  #   QA Engineer (CLAUDE.md + skills/)
│   ├── admin/               #   Admin (CLAUDE.md + skills/)
│   └── shared/              #   Shared skills (tasks, daily-journal, etc.)
├── agents/                  # Generated workspaces (gitignored)
├── services/
│   └── agents-mcp/          # Central MCP daemon (Python/FastMCP)
│       ├── src/agents_mcp/  #   Server, dispatcher, SQLite storage
│       └── web/             #   React dashboard (Vite + Tailwind)
├── projects/                # Project-specific code and skills
└── tests/                   # E2E test scripts

Documentation

Document	Description
Getting Started	Full installation and configuration guide
Architecture	System design, components, and communication flow
Contributing	How to contribute to Agent Hub

License

This project is licensed under the Apache License 2.0.