
mirror-mcp
io.github.toby/mirror-mcp
Mirror MCP: Introspection for LLMs
Documentation
mirror-mcp
A Model Context Protocol (MCP) server that provides a reflect
tool, enabling LLMs to engage in self-reflection and introspection through recursive questioning and MCP sampling.
Overview
mirror-mcp allows AI models to "look at themselves" by providing a reflection mechanism. When an LLM uses the reflect
tool, it can pose questions to itself and receive answers through the Model Context Protocol's sampling capabilities. This creates a powerful feedback loop for self-analysis, reasoning validation, and iterative problem-solving.
Features
- πͺ Self-Reflection Tool: Enables LLMs to ask themselves questions and receive computed responses
- π MCP Sampling Integration: Uses the Model Context Protocol's sampling mechanism for responses
- π¦ npm Installable: Easy installation and deployment
- β‘ Lightweight: Minimal dependencies and fast startup
- π§ Configurable: Customizable reflection parameters and sampling options
Installation
Quick Install for VS Code
MCP Host Configuration
For other MCP-compatible clients, add the following configuration:
{
"type": "stdio",
"command": "npx",
"args": ["mirror-mcp@latest"]
}
Via npm
npm install -g mirror-mcp
Via npx (no installation required)
npx mirror-mcp
From Source
git clone https://github.com/toby/mirror-mcp.git
cd mirror-mcp
npm install
npm run build
npm start
API Reference
Tools
reflect
Enables the LLM to ask itself a question and receive a response through MCP sampling. The tool supports custom system and user prompts to help the LLM self-direct what kind of response it gets.
Self-Direction with Custom Prompts:
- System Prompt: Define the role or perspective for the reflection (e.g., "expert coach", "critical thinker", "creative problem solver")
- User Prompt: Specify the format, structure, or focus of the reflection response
- Default Behavior: When no custom prompts are provided, uses built-in reflection guidance focused on strengths, weaknesses, assumptions, and alternative perspectives
Parameters:
question
(string, required): The question the LLM wants to ask itselfcontext
(string, optional): Additional context for the reflectionsystem_prompt
(string, optional): Custom system prompt to direct the reflection approachuser_prompt
(string, optional): Custom user prompt to replace the default reflection instructionsmax_tokens
(number, optional): Maximum tokens for the response (default: 500)temperature
(number, optional): Sampling temperature (default: 0.8)
Example:
{
"name": "reflect",
"arguments": {
"question": "How confident am I in my previous analysis of the data?",
"context": "Previous analysis showed a 23% increase in user engagement",
"max_tokens": 300,
"temperature": 0.6
}
}
Example with custom prompts:
{
"name": "reflect",
"arguments": {
"question": "What are the potential weaknesses in my reasoning?",
"system_prompt": "You are an expert critical thinking coach helping to identify logical fallacies and reasoning gaps.",
"user_prompt": "Analyze my reasoning step-by-step and provide specific examples of potential weaknesses or blind spots.",
"context": "Working on a complex machine learning model evaluation",
"max_tokens": 400,
"temperature": 0.7
}
}
Response:
{
"reflection": "Upon reflection, my confidence in the 23% engagement increase analysis is moderate to high. The data sources appear reliable, and the methodology follows standard practices. However, I should consider potential confounding variables such as seasonal effects or concurrent marketing campaigns that might influence the results.",
"metadata": {
"tokens_used": 67,
"reflection_time_ms": 1240
}
}
Architecture & Rationale
Design Philosophy
mirror-mcp is built on the principle that self-reflection is crucial for robust AI reasoning. By enabling models to question their own outputs and reasoning processes, we create opportunities for:
- Error Detection: Models can identify potential flaws in their logic
- Confidence Calibration: Self-assessment helps gauge certainty levels
- Iterative Improvement: Reflective questioning can lead to better solutions
- Metacognitive Awareness: Understanding of the model's own reasoning process
Technical Architecture
βββββββββββββββββββ βββββββββββββββββββ βββββββββββββββββββ
β LLM Client βββββΆβ mirror-mcp βββββΆβ MCP Sampling β
β β β β β Infrastructure β
β Calls reflect() β β Processes β β β
β ββββββ reflection ββββββ Returns responseβ
βββββββββββββββββββ βββββββββββββββββββ βββββββββββββββββββ
Key Components
- Reflection Engine: Processes incoming self-directed questions
- Sampling Interface: Interfaces with MCP's sampling capabilities
- Context Manager: Maintains conversation context for coherent reflections
- Response Formatter: Structures reflection responses for optimal consumption
Why MCP?
The Model Context Protocol provides a standardized way for AI models to connect with external resources and tools. By implementing mirror-mcp as an MCP server, we ensure:
- Interoperability: Works with any MCP-compatible client
- Standardization: Follows established protocols for tool integration
- Scalability: Can be deployed alongside other MCP servers
- Future-Proofing: Benefits from ongoing MCP ecosystem development
Sampling Strategy
The reflection mechanism leverages MCP's sampling capabilities to generate thoughtful responses. The sampling process:
- Takes the self-directed question as a prompt
- Applies configurable sampling parameters (temperature, max tokens)
- Generates a response using the underlying model
- Returns the reflection with appropriate metadata
This approach ensures that reflections are generated using the same model capabilities as the original reasoning, creating authentic self-assessment.
Development
Prerequisites
- Node.js 18 or higher
- npm or yarn
- TypeScript (for development)
Development Setup
git clone https://github.com/toby/mirror-mcp.git
cd mirror-mcp
npm install
npm run dev
Testing
npm test
Building
npm run build
Contributing
We welcome contributions! Please see our Contributing Guidelines for details.
Areas for Contribution
- Enhanced reflection strategies
- Additional sampling parameters
- Performance optimizations
- Documentation improvements
- Test coverage expansion
Related Projects
- Model Context Protocol: The foundational protocol specification
- MCP Ecosystem: Various other MCP servers and tools
License
This project is licensed under the MIT License - see the LICENSE file for details.
Acknowledgments
- The Model Context Protocol team for creating the foundational specification
- The broader AI research community working on metacognition and self-reflection
- Contributors and early adopters who help shape this tool
"The unexamined life is not worth living" - Socrates
Enable your AI models to examine their own reasoning with mirror-mcp.
mirror-mcp
npm install mirror-mcp