Token导航 LogoToken导航TokenDH.com
Creativity Generation Unit logo
办公协作未说明官方级别未说明来源级核验

Creativity Generation Unit

MCP Server

基于快思慢想理论的代理间创意生成服务,提供多层次的创意方法和工具,适用于需要创新解决方案的场景。

工具数

12

提示词数

0

GitHub Stars

0

资源数

0
设计PythonClaudeClaudeVS Code

安装说明

本站只整理中文说明和来源信息,不托管安装包,也不代用户安装。

作者 / 组织

u9401066

提供方

u9401066

最后核验

2026/5/17 20:22

快速接入

先看主来源和安装命令,再打开仓库或文档;下面只保留这个条目的关键接入事实。

详细介绍

Creativity Generation Unit (CGU)

🎨 MCP-based Agent-to-Agent Creative Idea Generator 基於快思慢想 (Thinking, Fast and Slow) 的創意發想服務

![License](https://opensource.org/licenses/Apache-2.0) ![Python](https://www.python.org/)

🌐 繁體中文

💡 Core Insight

"All models are wrong, but some are useful!" — George Box

Key Discovery: Creativity can emerge from partial information!

  • Humans don't need complete world knowledge to generate creative ideas
  • Creativity requires connection ability, not information volume
  • Even the simplest models can provide unique creative perspectives

🏗️ Architecture

┌──────────────────────────────────────────────────────────────┐
│               Creativity Generation Unit (CGU)                │
│                        MCP Server                             │
├──────────────────────────────────────────────────────────────┤
│                                                               │
│   ┌─────────────────────────────────────────────────────┐   │
│   │       🧠 v0.4.0: Agent-Driven Creativity                │   │
│   │                                                       │   │
│   │   ┌─────────────────────────────────────────────┐   │   │
│   │   │           Agent Creativity Toolbox              │   │   │
│   │   │  ┌─────────┐ ┌────────┐ ┌─────────┐ ┌──────┐ │   │   │
│   │   │  │ Concept │ │Connect-│ │ Novelty │ │ Idea │ │   │   │
│   │   │  │Explorer│ │  ion   │ │ Checker │ │Evolver│ │   │   │
│   │   │  │  🔍   │ │ Finder │ │   ✅    │ │  🧬  │ │   │   │
│   │   │  └─────────┘ └────────┘ └─────────┘ └──────┘ │   │   │
│   │   └─────────────────────────────────────────────┘   │   │
│   │                         │                               │   │
│   │   Agent decides:  ┌─────┴──────┐                      │   │
│   │   - Which tool?   │   Agent   │  ← Autonomous        │   │
│   │   - How to use?   │  Decision │    Exploration       │   │
│   │   - When to stop? └────────────┘                      │   │
│   └─────────────────────────────────────────────────────┘   │
│                                                               │
│   ┌─────────────────────────────────────────────────────┐   │
│   │        v0.2-v0.3: Core Engines (Foundation)           │   │
│   │  ┌─────────┐ ┌─────────┐ ┌──────────┐ ┌─────────┐ │   │
│   │  │ Analogy │ │  Graph  │ │Adversari-│ │Thinking │ │   │
│   │  │ Engine  │ │Traversal│ │al Engine │ │ Engine  │ │   │
│   │  └─────────┘ └─────────┘ └──────────┘ └─────────┘ │   │
│   └─────────────────────────────────────────────────────┘   │
│                                                               │
│   Backend: Ollama (Local) / Copilot (Framework Mode)         │
└──────────────────────────────────────────────────────────────┘

💡 Core Insight: Agent-Driven Creativity (v0.4.0)

"Copilot 內部觸碰不到,無論外層做什麼最終都是 Prompt 進去"

Key Shift: From "Human-Agent Language Interaction" to "Agent Autonomous Tool Interaction"

TraditionalAgent-Driven
We design the processAgent designs its own process
Fixed methodologyDynamic exploration strategy
Output cannot be verifiedTools can verify
One-shot generationIterative exploration

🧠 Thinking, Fast and Slow

Based on Daniel Kahneman's theory:

SystemSpeedCharacteristicsCGU Implementation
System 1Fast ⚡Intuitive, automaticREACT, ASSOCIATE, PATTERN_MATCH
System 2Slow 🐢Deliberate, analyticalANALYZE, SYNTHESIZE, EVALUATE
CreativeMixed 🎨Breaking boundariesDIVERGE, CONVERGE, TRANSFORM

Core Strategy: Multiple fast steps + occasional slow steps = Efficient creativity

🎯 Creativity Levels

Level 1: Combinational (0.7-1.0 association)
└─ New combinations of known elements

Level 2: Exploratory (0.3-0.7 association)
└─ Exploring boundaries within existing rules

Level 3: Transformational (0.0-0.3 association)
└─ Breaking rules, creating new paradigms

📚 15 Human Creativity Methods

CGU implements structured creativity methods:

CategoryMethods
DivergentMind Map, Brainstorm, SCAMPER, Random Input
Structural9-Grid Mandala, Morphological Analysis, 5W2H, Fishbone
PerspectiveSix Thinking Hats, Reverse Thinking, Analogy
ProcessDouble Diamond, Design Sprint, KJ Method, World Café
SystematicTRIZ 40 Principles

🛠️ Tech Stack

  • MCP SDK: FastMCP for tool serving
  • Agent Orchestration: LangGraph
  • Local Inference: vLLM + Qwen 4B
  • Structured Output: Pydantic + Instructor
  • Web Search: DuckDuckGo Search

🚀 Quick Start

# Clone repository
git clone https://github.com/YOUR_USERNAME/creativity-generation-unit.git
cd creativity-generation-unit

# Setup environment (uv recommended)
uv venv
uv sync --all-extras

# Run MCP server
cgu-server

# Or use CLI
cgu generate "How to improve remote work productivity?"

📁 Project Structure

creativity-generation-unit/
├── src/cgu/
│   ├── core/           # Core engines (v2)
│   │   ├── analogy.py  # Cross-domain analogy
│   │   ├── graph.py    # Concept graph traversal
│   │   ├── adversarial.py # Adversarial evolution
│   │   └── creativity_core.py # Unified engine
│   ├── tools/          # Agent tools (v0.4)
│   │   └── creativity_tools.py # 5 creativity tools
│   ├── soup/           # Spark-Soup 創意湯 (v0.5 NEW) 🆕
│   │   └── spark_soup.py # Context Stuffing for Creativity
│   ├── agents/         # Multi-Agent system (v0.3)
│   ├── thinking/       # Thinking Engine (v0.3)
│   ├── graph/          # LangGraph definitions
│   ├── llm/            # LLM backends
│   └── server.py       # MCP Server
├── docs/               # Documentation
├── tests/              # Test suite
├── memory-bank/        # Project memory
└── pyproject.toml      # Dependencies

🔧 MCP Tools

// Core Tools (v0.2)
generateIdeas(topic, creativityLevel, count)
sparkCollision(conceptA, conceptB)
associativeExpansion(seed, direction, depth)
applyMethod(method, input)

// Deep Thinking Tools (v0.3)
deepThink(topic, depth, mode)
multiAgentBrainstorm(topic, agents)
sparkCollisionDeep(conceptA, conceptB)

// Agent Creativity Tools (v0.4)
exploreConcept(concept)       // Search concept space
findConnection(a, b)          // Discover connections
checkNovelty(idea)            // Validate novelty
evolveIdea(idea, mutation)    // Mutate ideas
getProgress()                 // Track exploration

// Spark-Soup Tools (v0.5 NEW) 🆕
sparkSoupGenerate(topic)      // Assemble "creativity soup"
sparkSoupQuick(topic)         // Quick soup + idea generation
collectFragments(topic)       // Collect info fragments
getTriggerWords(categories)   // Get creativity triggers

🎮 Agent-Driven Creativity (v0.4.0)

Agent autonomously uses tools to explore creativity:

from cgu.tools import CreativityToolbox

toolbox = CreativityToolbox()

# Agent starts exploration
session = toolbox.start_session("remote work")

# Agent decides to explore concept
explore = toolbox.explore_concept("remote work")
# -> related: ['collaboration', 'efficiency', 'flexibility']
# -> unexpected: ['nomad', 'ritual', 'cafe']

# Agent tries cross-domain connection
connection = toolbox.find_connection("remote work", "nomad")
# -> novelty_score: 0.80

# Agent generates idea
idea = "Combine remote work with nomad lifestyle"
novelty = toolbox.check_novelty(idea)
# -> is_novel: True, score: 1.0

# If not novel, agent evolves
evolved = toolbox.evolve_idea(idea, "combine")

🌟 Design Principles

  1. Model Democracy - Even simple models have unique perspectives
  2. Partial is Enough - No need for complete world model
  3. Connection > Knowledge - Creativity is about linking
  4. Errors are Useful - Wrong connections may be innovations

🤖 OpenClaw Integration

CGU works natively with OpenClaw as an MCP tool server.

Passthrough Mode (Recommended for OpenClaw)

When running inside OpenClaw, your agents (Claude, GPT, etc.) are the LLM — no need for a secondary Ollama model. Use passthrough mode to get rich methodology frameworks that your agents fill with their own reasoning:

CGU_LLM_PROVIDER=passthrough  # Returns structured frameworks, no LLM call

What passthrough returns:

  • SCAMPER: all 7 dimensions with thinking angles and prompts
  • Six Hats: 6 perspectives with focus areas and guiding questions
  • Brainstorm: 3-round structure (wild → build → ground)
  • Every method includes _meta with instructions

OpenClaw Config

Add to your OpenClaw config.yaml:

mcp:
  servers:
    cgu:
      url: "http://localhost:8818/mcp"  # or your CGU server URL

Or run as stdio:

mcp:
  servers:
    cgu:
      command: "uv"
      args: ["--directory", "/path/to/creativity-generation-unit", "run", "cgu-server"]
      env:
        CGU_LLM_PROVIDER: "passthrough"

Agent-to-Agent Brainstorming

Use brainstorm_protocol to generate structured discussion scripts for two agents:

Agent A (domain expert) + Agent B (architect)
    │
    ▼
brainstorm_protocol(topic="...", method="six_hats")
    │
    ▼
Phase 1: Diverge → each agent explores from their angle
Phase 2: Collide → agents challenge each other's ideas  
Phase 3: Converge → jointly select best ideas
    │
    ▼
evaluate_brainstorm_ideas(ideas=[...])
    │
    ▼
Ranked results with feasibility/novelty/impact scores

Provider Modes

ModeLLMUse Case
ollamaLocal Ollama modelStandalone / offline use
passthroughNone (framework only)OpenClaw / any LLM-capable agent
copilot*(deprecated, alias for passthrough)*Legacy VS Code Copilot

📋 Documentation

📄 License

Apache License 2.0


*"Creativity is just connecting things."* — Steve Jobs

目录标签

目录标签

设计PythonClaude创意生成本地部署代理协作多方法融合创新工具思维模型

支持客户端

ClaudeVS Code

接入字段

传输方式(transport,传输协议)

未说明

鉴权方式(authType,认证方式)

session

工具数量(toolCount,工具数)

12

资源数量(resourceCount,资源数)

0

提示词数量(promptCount,提示词数)

0

权限和风险

未说明session部署方式未说明

接入前请确认传输方式、认证方式和部署位置,并根据实际工具能力限制访问范围。

安装前确认

不要直接授予不必要的文件、网络或账号权限;先核对安装命令和配置内容。

仍需确认:installCommand

来源信息

继续浏览同类 MCP