Created by 菠萝菠菠 🍍
English | 中文 | 🏛️ 当皇上 danghuangshang (OpenClaw版) | 🌐 官网 danghuangshang.com
TianGong is a self-evolving multi-agent AI framework that maps China's ancient imperial bureaucracy onto AI agent architecture. Instead of flat tool lists, agents are organized into a hierarchical government with planning, review, and execution layers — just like a real imperial court.
What makes it different from Claude Code / Cursor / Aider / OpenHands?
- 3 governance systems: Ming Dynasty cabinet, Tang Dynasty three-department, or Modern corporate — each with different agent counts, permission models, and review flows
- Self-evolving AI: Agents automatically optimize their own prompts based on performance data (the AGI loop)
- OpenViking memory: L0/L1/L2 tiered context loading that saves 80%+ tokens
- Agent competition: Pit agents against each other, hold court debates, run imperial exams
- Ancient wisdom injection: Classical Chinese philosophy woven into agent prompts
- MBTI × Zodiac personality: Each agent has unique personality traits that affect behavior
- Viral treasure hunt: A gamified system that drives organic adoption
git clone https://github.com/wanikua/tiangong.git
cd tiangong && npm install && npm link
tiangong setup # First-time config: choose provider + API key
tiangong # Enter interactive modeUpdate to latest version:
cd ~/Documents/tiangong && git pull && npm linkPrerequisites: Node.js >= 22
After npm link, just type tiangong in any terminal to enter interactive mode. Type your prompt directly — no flags needed, just like Claude Code.
No commands to memorize! Natural language works too:
> Have bingbu and gongbu compete on sorting ← auto-triggers /pk
> Let's discuss PostgreSQL vs MongoDB ← auto-triggers /debate
> Test bingbu's skills ← auto-triggers /exam
> Everyone work together on an auth module ← auto-triggers /collab
> Show me the rankings ← auto-triggers /rank
> How much did it cost ← auto-triggers /cost
# Install Ollama: https://ollama.com
# Then pull a model:
ollama pull qwen2.5-coder:7b
# Start tiangong, choose "Ollama" during setup
node bin/tiangong.js# Get API key: https://console.anthropic.com/settings/keys
# Start tiangong, choose "Anthropic", paste key
node bin/tiangong.js# Interactive mode (REPL) — recommended, type /help for all commands
node bin/tiangong.js
# Single command mode
node bin/tiangong.js "Build a login page"
# Choose governance system
node bin/tiangong.js --regime tang "Review this code for security"
node bin/tiangong.js --regime modern "Create a market analysis"
# Use specific provider/model
node bin/tiangong.js --provider ollama --model qwen2.5-coder:7b "Write a sort function"
# Dry-run (show plan without executing)
node bin/tiangong.js --dry-run "Refactor the auth module"
# Reconfigure
node bin/tiangong.js setup
# View court structure
node bin/tiangong.js court
# Global install (optional)
npm link
tiangong "Hello world"Type these in the interactive prompt:
/help — Show all commands
/court — View court hierarchy
/regime tang — Switch to Tang dynasty system
/cost — View token spending
/history — View past commands
/dream — AI predicts what you need next
/collab "task" — Multi-agent collaborative coding
/oracle <error> — Paste error log → auto fix
/pk a1 a2 "task" — Agent competition
/debate "topic" — Multi-agent court debate
/exam agentId — Benchmark an agent
/rank — XP leaderboard
/auto-optimize — Auto-optimize agent prompts
/viking ls — Browse OpenViking memory filesystem
/personality — View MBTI × Zodiac profiles
/treasure hunt — Play the treasure game
/exit — Exit
| Provider | Models | Local? |
|---|---|---|
| Anthropic | Claude Sonnet/Opus/Haiku | No |
| OpenRouter | 50+ models | No |
| OpenAI | GPT-4o, o3 | No |
| DeepSeek | Chat, Reasoner | No |
| Qwen | Max, Plus, Turbo | No |
| Ollama | Llama, Qwen, Mistral, etc. | Yes |
| LM Studio | Any GGUF model | Yes |
| UncommonRoute | Auto-select best model | Yes |
Ming Dynasty (Fast Iteration)
User ──→ 司礼监(Dispatch) ──→ 内阁(Optimize) ──→ 六部(Execute) ──→ 都察院(Review)
Tang Dynasty (Checks & Balances)
User ──→ 中书省(Draft) ──→ 门下省(Review) ──→ 尚书省(Execute) ──→ 给事中(Audit)
↑ │
└──────────── Veto & Return (封驳) ────────────────────────┘
Modern Corporate (Flat & Fast)
User ──→ CEO(Strategy) ──→ CTO/CFO/CMO(Division) ──→ Teams(Execute)
| Command | Feature | What it does |
|---|---|---|
/auto-optimize |
Self-Evolving Prompts | Agents analyze their own performance and rewrite their prompts |
/viking |
OpenViking Memory | L0/L1/L2 tiered context filesystem (viking://agent/, viking://user/) |
/dream |
Predictive Engine | Scans TODOs, git state, dependencies — predicts what you need next |
/collab |
Multi-Agent Coding | Architect + Coder + Security + Tester + Reviewer work simultaneously |
/oracle |
Crash Oracle | Paste any error log → auto-analysis + fix code generation |
| Command | Feature | What it does |
|---|---|---|
/pk |
Agent Arena | Two agents compete on the same task, judge picks winner |
/debate |
Court Debate | Multiple agents argue from different perspectives, generate consensus |
/exam |
Imperial Exam | 5-subject benchmark test (knowledge, security, architecture, writing, algorithms) |
/rank |
Reputation System | 19-rank progression from 从九品 to 太师, XP-based |
/personality |
MBTI × Zodiac | Each agent has personality traits that affect their responses |
/treasure |
Treasure Hunt | Gamified viral loop — find hidden treasures, share for bonuses |
| Command | Feature | What it does |
|---|---|---|
/replay |
Session Replay | Time-travel through past sessions, auto-generate weekly reports |
/autopsy |
Failure Analysis | Root cause analysis with cascade detection |
/evolve |
Regime Evolution | Auto-recommend the best governance system for your project |
/evolve-self |
Self-Evolution | Agent analyzes and improves its own code |
/cost |
Cost Tracking | Per-agent token usage and budget visualization |
/regime |
Switch Regime | Hot-swap governance systems mid-session |
/history |
Session History | Browse past commands and results |
Each agent's System Prompt is built from 6 layers:
┌─────────────────────────────────┐
│ 1. Role Definition (制度+职能) │ ← regimes.js
│ 2. Thinking Frameworks (思维框架) │ ← wisdom.js (MECE, 5 Whys, etc.)
│ 3. Ancient Wisdom (古文慧根) │ ← wisdom.js (22 classical quotes)
│ 4. MBTI × Zodiac (性格特质) │ ← agent-personality.js
│ 5. Auto-Optimized Overlay (自进化) │ ← auto-prompt-optimizer.js
│ 6. Viking Context (记忆上下文) │ ← viking-store.js (L0/L1/L2)
└─────────────────────────────────┘
tiangong/
├── bin/tiangong.js # CLI entry point
├── src/
│ ├── zhongshu/ # 中书省 — Planning Layer
│ │ ├── planner.js # Intent analysis + execution plan
│ │ ├── prompt-builder.js # 6-layer dynamic System Prompt
│ │ └── wisdom.js # Thinking frameworks + ancient wisdom
│ ├── menxia/ # 门下省 — Review Layer
│ │ ├── permission-gate.js # Permission matrix (wildcard support)
│ │ └── security-check.js # 30+ dangerous command patterns
│ ├── shangshu/ # 尚书省 — Execution Layer
│ │ ├── dispatcher.js # Agent loop (Anthropic + OpenAI format)
│ │ ├── bing/ # 兵部 — Bash + File I/O + Glob + Grep
│ │ ├── hu/ # 户部 — Cost tracking (10+ model pricing)
│ │ └── li/ # 礼部 — LLM API client (retry + fallback)
│ ├── engine/ # Core engine
│ │ ├── query-loop.js # Session orchestration with spinners
│ │ ├── repl.js # Interactive REPL (20+ commands)
│ │ └── spinner.js # Terminal animations
│ ├── memory/ # Memory systems
│ │ ├── store.js # Legacy memory store
│ │ ├── viking-store.js # OpenViking context filesystem
│ │ └── extractor.js # Auto memory extraction
│ ├── features/ # Unique features
│ │ ├── auto-prompt-optimizer.js # AGI self-evolution loop
│ │ ├── pk-arena.js # Agent competition
│ │ ├── court-debate.js # Multi-agent debate
│ │ ├── collaborative-coding.js # Parallel multi-role coding
│ │ ├── crash-oracle.js # Error log → fix generation
│ │ ├── dream-engine.js # Predictive TODO/git/dep analysis
│ │ ├── imperial-exam.js # Agent benchmark testing
│ │ ├── reputation.js # XP + 19-rank system
│ │ ├── agent-personality.js # MBTI × Zodiac
│ │ ├── treasure-hunt.js # Viral gamification
│ │ ├── time-travel.js # Session replay + weekly reports
│ │ ├── autopsy.js # Failure root cause analysis
│ │ ├── regime-evolution.js # Smart regime recommendation
│ │ └── self-evolution.js # Self-improvement engine
│ ├── export/ # OpenClaw export
│ └── config/ # Providers + Regimes + Setup
TianGong is an independent, original implementation. All code is original work under the BSL 1.1 License. The following projects provided architectural inspiration:
- 当皇上 (danghuangshang) — AI imperial court tutorial + OpenClaw implementation. TianGong's regime system and agent hierarchy concepts originated here.
- Claude Code Architecture — Anthropic's CLI tool architecture study. Inspired our tool system design (Bash/FileRead/FileEdit/Glob/Grep), permission model, and agent loop pattern.
- OpenViking — ByteDance's open-source context database for AI Agents. Our
viking-store.jsimplements the L0/L1/L2 tiered context loading andviking://filesystem paradigm described in their paper. - Thinking Frameworks — Structured thinking models. Our
wisdom.jsembeds 8 frameworks (MECE, 5 Whys, First Principles, etc.) into agent prompts.
- Thinking Frameworks — Structured thinking models collection
- danghuangshang — AI imperial court tutorial
天工开物是一个自进化多 Agent AI 框架，将中国古代三省六部的朝廷制度映射到 AI Agent 架构上。Agent 不再是扁平的工具列表，而是按决策层、审核层、执行层组织的朝廷官僚体系。
和 Claude Code / Cursor / Aider / OpenHands 有什么不同？
- 三种治国方略：明朝内阁制（快速迭代）、唐朝三省制（三权制衡+封驳权）、现代企业制（扁平高效）
- AI 自进化：Agent 根据历史表现数据，自动重写自己的 Prompt（AGI 闭环）
- OpenViking 记忆：L0/L1/L2 三层上下文按需加载，节省 80%+ token
- Agent 对决：武举殿试、廷议辩论、科举考试——大臣们互相竞争、辩论、考核
- 古文慧根植入：22 条经典古文智慧融入 Agent Prompt（「凡事预则立，不预则废」）
- MBTI × 星座性格：每个大臣有独特性格，影响回复风格和合拍度
- 寻宝游戏：病毒式传播机制——宝藏收集、谜语挑战、邀请码裂变
git clone https://github.com/wanikua/tiangong.git
cd tiangong && npm install && npm link
tiangong setup # 首次配置：选择 Provider + API Key
tiangong # 进入交互模式更新到最新版：
cd ~/Documents/tiangong && git pull && npm link前置条件: Node.js >= 22
npm link 之后，任意终端输入 tiangong 即进入交互模式。直接打字说话就行，像 Claude Code 一样简单。
不用记命令！ 自然语言也能触发功能：
> 让兵部和工部比一比写排序算法 ← 自动触发 /pk
> 大家讨论一下用 PostgreSQL 还是 MongoDB ← 自动触发 /debate
> 考一考兵部 ← 自动触发 /exam
> 大家一起写一个认证模块 ← 自动触发 /collab
> 看看排行榜 ← 自动触发 /rank
> 花了多少钱 ← 自动触发 /cost
> 帮助 ← 自动触发 /help
# 安装 Ollama: https://ollama.com
# 拉取模型:
ollama pull qwen2.5-coder:7b
# 启动天工，选择 "Ollama" 即可
node bin/tiangong.js# 获取 API Key: https://console.anthropic.com/settings/keys
# 启动后选择 "Anthropic"，粘贴 Key
node bin/tiangong.js# 交互模式（推荐），输入 /help 查看所有命令
node bin/tiangong.js
# 单次执行模式
node bin/tiangong.js "帮我写一个登录页面"
# 选择制度
node bin/tiangong.js --regime tang "审查这段代码"
node bin/tiangong.js --regime modern "做一个市场分析"
# 指定模型
node bin/tiangong.js --provider ollama --model qwen2.5-coder:7b "写排序算法"
# 试运行（只看计划不执行）
node bin/tiangong.js --dry-run "重构认证模块"
# 重新配置
node bin/tiangong.js setup
# 全局安装（可选）
npm link
tiangong "你好世界"在交互提示符中输入：
/help 帮助
/court 查看朝廷架构
/regime tang 切换到唐制
/cost 查看花费
/history 查看历史
/dream AI 预判你下一步需求
/collab "任务" 多 Agent 协同编码
/oracle <错误日志> 粘贴错误自动修复
/pk a1 a2 "题目" Agent 对决
/debate "议题" 廷议辩论
/exam bingbu 科举考试
/rank 功勋排行榜
/auto-optimize 自动优化 Agent Prompt
/viking ls 浏览 OpenViking 记忆
/personality 查看 MBTI 性格档案
/treasure hunt 寻宝游戏
/exit 退朝
| 命令 | 功能 | 说明 |
|---|---|---|
/auto-optimize |
🧬 自动 Prompt 优化 | AGI 核心——Agent 分析自己的表现并重写 Prompt |
/viking |
📂 OpenViking 记忆 | L0/L1/L2 三层上下文文件系统 |
/dream |
🔮 朝堂梦境 | 扫描 TODO/git/依赖，预判你下一步需求 |
/collab |
📋 六部联名 | 架构师+编码者+安全官+测试官+审查官协同 |
/oracle |
📜 天书降世 | 粘贴错误日志，自动根因分析+修复代码 |
/pk |
⚔️ 武举殿试 | Agent 对决擂台 |
/debate |
📣 廷议 | 多 Agent 朝堂辩论，自动总结共识和争议 |
/exam |
📝 科举考试 | 5 科能力基准测试（明经/明法/策论/诗赋/算术） |
/rank |
🏆 功勋排行 | 19 级品阶（从九品→太师），XP 驱动 |
/personality |
🧬 性格档案 | 16 种 MBTI × 12 星座 × 合拍度分析 |
/treasure |
🗺️ 寻宝奇缘 | 宝藏收集+谜语+邀请码裂变 |
/replay |
📜 奏折回放 | 会话时间旅行+自动周报 |
/autopsy |
🔍 大理寺 | 故障验尸报告+级联效应分析 |
/evolve |
👑 朝代更迭 | 智能制度自适应推荐 |
/evolve-self |
🧬 自进化 | Agent 自我改进引擎 |
/cost |
💰 户部账目 | 预算进度条可视化 |
/regime |
制度切换 | 运行时热切换制度 |
/history |
旨意历史 | 浏览历史命令 |
角色定义 → 思维框架(8种) → 古文慧根(22条) → MBTI性格 → 自进化Overlay → Viking上下文
每一层独立进化，互不干扰。
BSL 1.1 (Business Source License) — 非商用免费，商用需授权。2030-04-02 后自动转为 Apache 2.0。
「思考，是人类最后的自由。」—— 查理·芒格
"The right thinking framework matters more than intelligence." — Charlie Munger
| 小红书「菠萝菠菠🍍」 | 公众号「菠言菠语」 | 微信群 |
|---|---|---|
v0.1.0 | BSL 1.1 License | User Agreement | Privacy Policy
📜 Licensed under BSL 1.1. Credit: danghuangshang by @wanikua