Mirror MCP (io.github.toby/mirror-mcp) is a Model Context Protocol (MCP) server that provides a reflect tool for LLM introspection. Its goal is to enable self-reflection by allowing recursive questioning and MCP sampling so models can request and receive answers about themselves.
🛠️ Key Features
Provides a reflect tool
Enables introspection via recursive questioning
Supports “MCP sampling” when using reflect
🚀 Use Cases
LLM self-reflection workflows using the reflect tool
Introspection tasks where models query themselves for answers
⚡ Developer Benefits
Standard MCP server interface with a dedicated reflect tool
Exposes an introspection mechanism (“look at themselves”) for LLMs
⚠️ Limitations
Functionality described is limited to the reflect tool and its reflection/introspection behavior
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
🪞 Self-Reflection Tool: Enables LLMs to ask themselves questions and receive computed responses
🔄 MCP Sampling Integration: Uses the Model Context Protocol's sampling mechanism for responses
📦 npm Installable: Easy installation and deployment
⚡ Lightweight: Minimal dependencies and fast startup
🔧 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:
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 itself
context (string, optional): Additional context for the reflection
system_prompt (string, optional): Custom system prompt to direct the reflection approach
user_prompt (string, optional): Custom user prompt to replace the default reflection instructions
max_tokens (number, optional): Maximum tokens for the response (default: 500)
temperature (number, optional): Sampling temperature (default: 0.8)
Example:
json
{"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:
json
{"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:
json
{"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
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
bash
git clone https://github.com/toby/mirror-mcp.git
cd mirror-mcp
npm install
npm run dev
