AI导哈

发散思维MCP服务器

为AI代理提供结构化发散思维工具的MCP服务器，支持简单模式与专业模式，帮助生成多样化创意并避免思维趋同。

  • by d4vidc4rson
  • TypeScript
  • 混合部署
  • 1 星标
  • 更新于
在 GitHub 查看 d4vidc4rson/divergent-agents

使用说明

Divergent Agents

Domain-specific AI agents powered by structured creativity tools. Stop getting the same ideas as everyone else.

This repo provides an MCP server with two modes:

  • Simple mode — one divergent_think tool that automatically routes to the right thinking technique. Drop it into any MCP client and go.
  • Pro mode — 14 modular tools (diagnose_problem, generate_diverse, audit_diversity, etc.) designed to be orchestrated by domain agent system prompts.

Quick start

Local (stdio)

cd mcp-server
npm install
npm run build

# Simple mode (default) — 1 tool
MODE=simple node dist/index.js

# Pro mode — 14 tools
MODE=pro node dist/index.js

# All tools — 15 tools (simple + pro)
MODE=all node dist/index.js

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "divergent-agents": {
      "command": "node",
      "args": ["/path/to/divergent-agents/mcp-server/dist/index.js"],
      "env": {
        "MODE": "simple"
      }
    }
  }
}

Remote (HTTP)

cd mcp-server
TRANSPORT=http PORT=3000 MODE=simple node dist/index.js

Connect any MCP client to http://your-host:3000/mcp.

Claude.ai (remote MCP)

Use the hosted URL as a remote MCP server in Claude.ai settings. Set MODE=simple for general use, or MODE=pro with a domain agent Project.

Modes

Mode Tools Use case
simple 1 (divergent_think) General use in any MCP client — the router picks the right technique
pro 14 modular tools Use with a domain agent system prompt in a Claude Project
all 15 (both) Development/testing

Pro mode tools

Tool What it does
diagnose_problem Check if you're solving the right problem
generate_diverse Generate genuinely different ideas using Guilford's framework
generate_personas Ideas from radically different constructed personas
think_wrong Push toward counterintuitive, non-obvious approaches
audit_diversity Score whether your ideas are actually different
map_blind_spots Find what you forgot to think about
extract_desire Strip briefs to raw human desire before ideating
generate_fast Rapid gut-level ideas, no deliberation
explore_range Ideas across the full moonshot-to-Monday spectrum
generate_bad Deliberately terrible ideas that unlock hidden value
generate_constrained Solutions using only resources already available
inject_random Genuine randomness to break exhausted patterns
ask_questions Generate questions, not answers
humanize_text Transform AI-generated text into human-sounding copy

Domain agents

Domain agents are system prompts that orchestrate pro-mode tools for specific roles. Paste them into Claude Project Instructions (or equivalent).

Available agents:

Using a domain agent

  1. Deploy or run the MCP server in pro mode
  2. Create a new Claude Project
  3. Add the MCP server as a remote tool
  4. Paste the agent system prompt (e.g., agents/product.md) into Project Instructions
  5. Start working — the agent knows when and how to use each tool

Deployment

Railway

# Push to Railway — uses railway.json for build config
railway up

Set environment variables:

  • TRANSPORT=http
  • MODE=simple (or pro)
  • PORT is set automatically by Railway

License

MIT

相关 MCP