MCP Dual-Cycle Reasoner

A Model Context Protocol (MCP) server implementing the Dual-Cycle Metacognitive Reasoning Framework for autonomous agents. This tool empowers agents with greater self-awareness and reliability through intelligent loop detection and experience acquisition.

Description

The MCP Dual-Cycle Reasoner is a sophisticated tool designed to enhance the autonomy and reliability of AI agents. By implementing a dual-cycle metacognitive framework, it provides agents with the ability to monitor their own cognitive processes, detect when they are stuck in repetitive loops, and learn from past experiences to make better decisions.

The framework consists of two main components:

• Sentinel: Monitors the agent's actions and detects anomalies, such as action repetition, state invariance, and progress stagnation.
• Adjudicator: Manages a case base of past experiences, allowing the agent to store and retrieve solutions to previously encountered problems.

This server is built with TypeScript and leverages high-performance libraries for statistical analysis, natural language processing, and semantic similarity, enabling advanced features like entropy-based anomaly detection, NLI-based text analysis, and intelligent case management.

Key Features

• 📊 Advanced Statistical Analysis: Entropy-based anomaly detection and time series analysis.
• 🧠 Enhanced Case-Based Reasoning: Semantic similarity matching with NLI-based text analysis.
• 🎯 Multi-Strategy Detection: Statistical, pattern-based, and hybrid loop detection.
• 📈 Time Series Analysis: Trend detection and cyclical pattern recognition.
• 🔧 Configurable Detection: Domain-specific thresholds and progress indicators.
• 🎨 Intelligent Case Management: Quality scoring, deduplication, and usage-based optimization.
• 🚀 High-Performance Libraries: Built with simple-statistics, natural, compromise, and HuggingFace Transformers.

Tech Stack

• Language: TypeScript
• Framework: Node.js
• Server: FastMCP for SSE transport
• NLP and Machine Learning:
  • @huggingface/transformers: For NLI-based semantic analysis
  • natural: For sentiment analysis and tokenization
  • compromise: For natural language processing
• Statistics:
  • simple-statistics: For statistical calculations
  • ml-matrix: For matrix operations
• Development Tools:
  • jest: For testing
  • eslint: For linting
  • prettier: For code formatting
  • zod: For schema validation

Installation

To get the project running locally, follow these steps:

1. Clone the repository:
   git clone https://github.com/cyqlelabs/mcp-dual-cycle-reasoner.git cd mcp-dual-cycle-reasoner
2. Install dependencies:
   npm install
3. Build the project:
   npm run build

Usage

Running the Server

You can run the server in two modes:

1. HTTP Stream (Default):
   npm start
   The server will start on port 8080.
2. Stdio:
   npm start -- --stdio

Using with Claude Desktop

Add the following to your Claude Desktop MCP configuration:

For stdio transport, add the --stdio flag to the args array.

Available Tools

Core Monitoring Tools

start_monitoring

Initialize metacognitive monitoring of an agent's cognitive process.

Input Schema:

process_trace_update

Main monitoring function—processes cognitive trace updates from the agent.

Input Schema:

Return Payload:

Returns a JSON object indicating if intervention is required and details about any detected loop.

stop_monitoring

Stop metacognitive monitoring and get a session summary.

Input Schema: {}

Loop Detection Tools

detect_loop

Detect if the agent is stuck in a loop using various strategies.

Input Schema:

Return Payload:

Returns a LoopDetectionResult object as a JSON string.

configure_detection

Configure loop detection parameters and domain-specific progress indicators.

Input Schema:

Enhanced Experience Management

store_experience

Store a case for future case-based reasoning with enhanced metadata and quality scoring.

Input Schema:

retrieve_similar_cases

Retrieve similar cases using advanced semantic matching and filtering.

Input Schema:

Return Payload:

Returns an array of Case objects as a JSON string.

System Tools

get_monitoring_status

Get the current monitoring status and statistics.

Input Schema: {}

Return Payload:

Returns a JSON string with the current monitoring status, including is_monitoring, current_goal, trace_length, and intervention_count.

reset_engine

Reset the dual-cycle engine state.

Input Schema: {}

Example Usage Scenario

Example Usage Scenario

Here's a complete example showing how to use the dual-cycle reasoner to monitor an autonomous agent and build up experience over time:

1. Initial Setup and Configuration

2. Monitoring Agent Actions

3. Storing and Retrieving Experiences

Understanding Loop Detection

When the process_trace_update tool detects a potential loop, it returns a detailed loop_detected object. Understanding the components of this object can help you diagnose and debug agent behavior.

Here's a quick guide to the key metrics in the details object:

• dominant_method: The primary method that triggered the loop detection (e.g., semantic_repetition, state_invariance).
• anomaly_score: An overall score (0-1) representing the confidence that the agent's recent actions are part of a loop. It's a weighted average of multiple detection methods.
• actions_involved_count: The number of recent actions identified as being part of the detected loop.
• recent_actions_count: The total number of recent actions analyzed.
• metrics: An object containing the raw scores from the different detection methods, such as:
  • semantic_repetition: The percentage of recent actions that are semantically similar to each other.
  • parameter_repetition: The percentage of similarity between the parameters of semantically similar actions.
  • exact_repetition: The percentage of recent actions that are exact, character-for-character duplicates.
  • cyclical_pattern: The score (0-1) indicating the presence of repeating sequences of actions.
  • oscillation_pattern: The score (0-1) indicating the presence of oscillating (back-and-forth) actions.
  • alternating_pattern: The score (0-1) indicating the presence of alternating action patterns.

Contributing

Contributions are welcome! Please read the contributing guidelines and ensure all tests pass before submitting a pull request.

License

This project is licensed under the MIT License. See the LICENSE file for details.