English | 繁體中文

AgentHub is not another AI chat interface.

It is a Harness Engineering system that lets you build a virtual development team with Claude Code — complete with a PM, Tech Lead, frontend developer, backend developer, and designer — and manage them like a real boss.

You give the orders. Agents execute. Hooks enforce quality. Skills standardize process. FileWatcher syncs in real time.

No prompt prayers. No "hope the AI remembers the rules." Rules live in Hooks. Agents that break them get blocked. It's that simple.

You Use Claude Code to Write Code — But Have You Ever Run Into This?

• Every new session wipes the lessons learned from the last one
• You tell the AI "don't force push," and it forgets again next time
• You spend 30 minutes crafting the perfect prompt, and the AI only follows the first three rules
• When multiple Agents work in parallel, no one knows who changed what

AgentHub doesn't solve "AI isn't smart enough." It solves "smart AI without discipline."

Why Is This Not "Just Another AI Wrapper"?

Most AI tools follow this logic: write better prompts → pray the AI complies.

AgentHub follows a different logic: replace text prayers with architectural constraints.

A good validator with a bad workflow beats a good workflow without a validator. This isn't a slogan — it's math: 5 steps at 80% success each = 33% overall. Add a validator that allows retries = 99%.

Core Architecture: Harness Engineering

Just like a race car needs more than a great engine — it needs seatbelts, guardrails, and a pit stop process.

Skill — Standardized Workflows

Reusable workflow templates that automatically load the relevant guidance when an Agent runs them.

• /sop-execute — L2 execution SOP: mandatory spec loading → implement → acceptance verification
• /sop-plan — L1 planning SOP: context loading → complexity analysis → Plan Mode decision → task breakdown
• /sop-review — L1 review SOP: load standards → code review → approve or reject
• /sop-deploy — Deploy SOP: quality checklist → pre-deploy → Gate G5
• /task-dispatch — Boss creates a task with one command, automatically written to the dev plan
• /review — Auto-detects the step and selects the appropriate review type
• /pre-deploy — Pre-deployment automated checks (CI / environment variables / Docker)
• /harness-audit — Periodic health scan scoring against seven core principles
• ...a total of 24 built-in Skills

Hook — Automated Interceptors

Not after-the-fact reminders — real-time blocking. Dangerous operations are stopped the moment they happen.

• PreToolUse: checks before command execution (blocks kill-port / --no-verify / force push main / malformed commit messages)
• PostToolUse: alerts after file modification (if a core service changes, forces .knowledge/ doc sync)
• Stop: validates before finishing (tests + type checks must pass, or the Agent cannot stop; also detects missing global Skills)

FileWatcher — Real-Time Sync

Markdown files are the database. Edit a .tasks/ file and the GUI updates instantly.

.tasks/T5.md is modified
    → chokidar detects the change
    → markdown-parser parses the file
    → DB upsert
    → eventBus broadcasts
    → Vue reactive update
    → GUI reflects the change instantly

Gate — Quality Checkpoints

G0 (Requirements Confirmation) → G1 (Design Review) → G2 (Code Review) → G3 (QA Acceptance) → G4 (Documentation Review) → G5 (Deploy Readiness) → G6 (Production Release)

You cannot proceed without passing. Not enforced by willpower — enforced by architecture.

What Does It Look Like?

An Electron desktop app with a dark theme.

Sessions — Agent Work Sessions

An embedded xterm.js terminal lets you interact with Claude Code Agents directly inside the GUI.

Task Board — Kanban-Style Task Management

A five-column kanban board with automatic status transitions. Click any card to see full task details.

Projects — Sub-Project Management

One-click Harness scaffolding with live display of task completion rate, active Sprint, and latest Gate status.

Agents — Team Member Overview

Every Agent has its own role definition, permission scope, and reporting chain.

Harness — Trigger Logs

Real-time Hook execution history with filtering, statistics, and rankings at a glance.

Team Structure

You are the Boss. Below you is your virtual software company — 9 departments, 46 Agents:

Chain of Command

Boss (You)
├── L1 Leadership (report directly to Boss)
│   ├── Product Manager        — Product strategy and roadmap
│   ├── Tech Lead              — Technical decisions and architecture
│   ├── Design Director        — UI/UX and design systems
│   ├── Marketing Lead         — Marketing strategy
│   ├── QA Lead                — Quality assurance
│   ├── Project Lead           — Project scheduling and milestones
│   ├── Operations Lead        — Studio operations
│   └── Company Manager        — Company-wide knowledge management
│
└── L2 Execution (report to L1, cannot escalate directly to Boss)

Full Department Roster

Product

Engineering

Design

Marketing

Testing

Project Management

Studio Operations

Bonus — Special Roles

Every Agent has its own role definition file (agents/definitions/), permission scope, and reporting chain. L2 cannot bypass L1 to reach the Boss, and the Boss cannot bypass L1 to directly command L2. Just like a real company.

Tech Stack

Prerequisites

AgentHub is a management framework built on top of Claude Code. All actual Agent work is executed through the Claude Code CLI.

Platform Setup

Windows

# Install Windows Build Tools (Admin PowerShell)
npm install --global windows-build-tools
# Or manually install Visual Studio Build Tools (select the "Desktop development with C++" workload)

macOS

# Install Xcode Command Line Tools
xcode-select --install

Linux (Ubuntu/Debian)

sudo apt-get install -y build-essential python3

Without Claude Code, AgentHub is just an empty GUI shell. All Skill execution, Hook interception, and Agent conversations depend on the Claude Code CLI.

Quick Start

# 1. Confirm prerequisites are installed
node --version    # >= 18
claude --version  # Claude Code CLI

# 2. Clone the project
git clone https://github.com/Stanshy/AgentHub.git
cd AgentHub

# 3. Install dependencies
npm install

# 4. Install Skills globally (one-time setup)
bash scripts/install-skills.sh

# 5. Start in development mode
npm run dev

Troubleshooting npm install failure: If you see Could not find any Visual Studio installation or node-gyp failed to rebuild node-pty, it means C++ Build Tools are not installed. See Platform Setup above. As a quick workaround, you can run npm install --ignore-scripts to skip native module compilation, then npm run dev — the terminal feature may not work, but the rest of the app will start.

Available Commands

npm run dev          # Start development mode (Electron + Vite HMR)
npm run build        # Build (TypeScript compilation + Vite build)
npm run typecheck    # TypeScript type checking
npm run test         # Unit tests (Vitest)
npm run test:e2e     # E2E tests (Playwright)
npm run build:win    # Build Windows installer
npm run build:mac    # Build macOS installer

Learning Resources

AgentHub's design philosophy is grounded in the Harness Engineering methodology from the Claude Code Mastery course.

If you want to understand the mental model behind AgentHub — why Hooks instead of prompts, why Skills are more effective than SOP documents, why Gate checkpoints can push a 33% success rate to 99% — this course is the best place to start:

Claude Code Mastery — The complete guide from zero to autonomous Agent teams Read Online — Web version for easy browsing

Covers 8 modules and 23 chapters, from Claude Code fundamentals to the five-stage evolution model (Manual Operation → Prompt Engineering → Harness Engineering → Autonomous Agents → Agent Teams). Module 6 is dedicated to the Harness Engineering architecture used by AgentHub.

You can use AgentHub standalone, or pair it with the course to understand the "why" behind every design decision.

Design Philosophy

A good validator with a bad workflow beats a good workflow without a validator.

1. Give a map, not an encyclopedia — CLAUDE.md is an index, not a full specification
2. Enforce with tools, not prompts — Forbidden commands live in Hooks, not documents
3. Knowledge compounds — Postmortem notes are automatically collected, fed back across projects, and inherited by new ones
4. Processes are repeatable — Skills standardize every step, removing reliance on personal memory
5. Quality is guaranteed — Gate checkpoints + Hook interception provide double-layer protection

What Can It Do Now? Where Is It Going?

Now (v0.1) — Implemented

Project Management

• Auto-scaffolds a complete Harness (CLAUDE.md + .knowledge/ + Skills + Hooks) when a sub-project is created
• Supports 4 project templates (web-app / api-service / library / mobile-app)
• Project cards display live task completion rate, active Sprint, and latest Gate status

Sessions & Terminal

• Embedded xterm.js terminal for opening Claude Code Sessions directly inside the GUI
• Agent definition files auto-load, bringing each Session the relevant Agent's role and standards
• Session history and token usage tracking

Task Board

• Five-column kanban: Created → Assigned → In Progress → In Review → Done
• Agents update status in sub-projects via /task-start, /task-done, and /task-approve Skills
• Task files follow Sprint subdirectory convention: .tasks/sprint-{N}/TN-xxx.md
• File changes → chokidar detects → markdown-parser parses → DB syncs → GUI updates live

Harness System

• 24 Skill templates: SOP enforcement (sop-plan / sop-execute / sop-review / sop-deploy), Sprint proposals, task dispatch, Code Review, Gate records, pre-deployment checks, and more
• 6 Hook templates: forbidden-commands (dangerous command blocking), git-commit-check (Conventional Commits validation), stop-validator (validates tests + type checks before stopping), g1/g4/g5 quality gate checks
• Skills are installed globally to ~/.claude/commands/ via scripts/install-skills.sh — no per-project duplication needed
• Hooks are automatically deployed to the sub-project's .claude/ directory on project creation
• The GUI lets you browse, add, and edit Hooks, with support for both global and project scopes

Gate Quality Checkpoints

• G0–G6 seven checkpoints; changes to section 10 of the dev-plan sync to the GUI in real time
• Pipeline view shows the review status of each checkpoint

Knowledge Base

• Browse the .knowledge/ directory tree with live Markdown preview
• Company standards (coding-standards, api-standards, testing-standards) managed centrally

Your workflow: Open a Session in the GUI → give orders to an Agent → the Agent works inside the sub-project (guided by Skills, constrained by Hooks) → file changes sync back to the GUI instantly → you review results and approve Gates.

Future (Roadmap)

• Automated Chain of Command: Boss issues one directive, PM automatically breaks it down and dispatches to TL and DD — no manual hand-off required
• Parallel Sessions: Run multiple Agent Sessions simultaneously with automatic task allocation and coordination
• Runtime Guardrails: TypeScript engine intercepts in real time — not just CLI commands, but code logic as well
• Cross-Project Knowledge Network: A pitfall discovered in Project A automatically becomes a protective rule for all new projects

You Don't Have to Do Everything. You Only Have to Do What a Boss Does.

You open your laptop and see the Dashboard showing — 3 Sprints in progress, 12 tasks assigned to different Agents, the Tech Lead just finished a Code Review, the PM is preparing a Gate review, and one Agent tried to skip the tests — the Stop Hook blocked it and it self-corrected.

Your job: make decisions, review outputs, set direction. Not your job: write prompt prayers, run tests manually, worry about anyone cutting corners.

That's AgentHub.

Acknowledgements

• Agent architecture design inspired by contains-studio/agents

License

MIT