wisdom layer mcp

A metacognitive advisor for Claude powered by LearnLM.

Overview

The Wisdom Layer MCP is a specialized Model Context Protocol server designed to enhance Claude's reasoning capabilities. It creates a "wisdom layer" that acts as a strategic questioner, advisor, and complexity reducer for Claude, helping it to avoid recurring mistakes and achieve better results.

Features

• Strategic Metacognitive Advisor: Analyzes Claude's thinking process and provides targeted advice
• Plan Distillation: Forces ultra-simplified summaries for clarity and focus
• Mistake Tracking: Learns from past mistakes to avoid future ones
• LearnLM Integration: Powered by Google's LearnLM/Gemini API for metacognitive guidance

Tools

1. wisdom_advise

A constraint-free strategic advisor that works like a challenging mentor:

• Takes raw, unfiltered context to avoid Claude's cognitive biases
• Provides user alignment checks and complexity reduction
• Offers pattern-breaking questions and perspectives
• Suggests appropriate MCP tools to use
• Learns from past mistake patterns

<use_mcp_tool>
<server_name>wisdom</server_name>
<tool_name>wisdom_advise</tool_name>
<arguments>
{
  "plan": "My current plan is...",
  "userRequest": "Original request from user",
  "thinkingLog": "Raw sequential thinking output",
  "availableTools": ["tool1", "tool2"]
}
</arguments>
</use_mcp_tool>

2. wisdom_canvas

A distillation tool for final plans:

• Forces ultra-simplified plan representation
• Includes "why" section to justify approach
• Creates a checkpoint before implementation
• Serves as a reference during execution

<use_mcp_tool>
<server_name>wisdom</server_name>
<tool_name>wisdom_canvas</tool_name>
<arguments>
{
  "plan": "Detailed plan to distill",
  "userRequest": "Original request from user"
}
</arguments>
</use_mcp_tool>

3. wisdom_log

A mistake tracking system:

• One-sentence descriptions of mistakes made and corrected
• Categorizes and tallies recurring mistakes
• Creates a persistent learning feedback loop
• Read by LearnLM to provide targeted guidance

<use_mcp_tool>
<server_name>wisdom</server_name>
<tool_name>wisdom_log</tool_name>
<arguments>
{
  "mistake": "One-sentence description of the mistake",
  "category": "mistake-category",
  "solution": "How it was corrected"
}
</arguments>
</use_mcp_tool>

Installation

Prerequisites

• Node.js (v18.0.0 or higher)
• npm (v7.0.0 or higher)
• A Gemini API key for LearnLM functionality

Installation Steps

# Install globally
npm install -g wisdom-layer-mcp

# Or run directly
npx wisdom-layer-mcp

MCP Configuration

Add the server to your MCP settings file. For Claude/Cline, this is typically located at: - For Claude Desktop: ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) - For VSCode Cline: ~/.config/Code/User/globalStorage/saoudrizwan.claude-dev/settings/cline_mcp_settings.json (Linux)

Add the following configuration:

{
  "mcpServers": {
    "wisdom": {
      "command": "wisdom-layer-mcp",
      "args": [],
      "env": {
        "GEMINI_API_KEY": "your-gemini-api-key"
      },
      "disabled": false,
      "autoApprove": []
    }
  }
}

Usage Pattern

The ideal usage pattern for Claude:

1. Claude uses sequential thinking to formulate an initial plan
2. Claude calls wisdom_advise with raw context
3. Claude refines its approach based on LearnLM's advice
4. Claude finalizes with wisdom_canvas for ultra-clarity
5. Claude implements the solution
6. Claude logs any lessons learned in wisdom_log

This creates a metacognitive layer that helps Claude think better about its own thinking.

Development

# Clone the repository
git clone https://github.com/yourusername/wisdom-layer-mcp.git
cd wisdom-layer-mcp

# Install dependencies
npm install

# Build
npm run build

# Start development mode
npm run dev

License

MIT