Recursive Companion MCP

An MCP (Model Context Protocol) server that implements iterative refinement through self-critique cycles. Inspired by Hank Besser's recursive-companion, this implementation adds incremental processing to avoid timeouts and enable progress visibility.

Features

• Incremental Refinement: Avoids timeouts by breaking refinement into discrete steps
• Mathematical Convergence: Uses cosine similarity to measure when refinement is complete
• Domain-Specific Optimization: Auto-detects and optimizes for technical, marketing, strategy, legal, and financial domains
• Progress Visibility: Each step returns immediately, allowing UI updates
• Parallel Sessions: Support for multiple concurrent refinement sessions

How It Works

The refinement process follows a Draft → Critique → Revise → Converge pattern:

1. Draft: Generate initial response
2. Critique: Create multiple parallel critiques (using faster models)
3. Revise: Synthesize critiques into improved version
4. Converge: Measure similarity and repeat until threshold reached

Installation

Prerequisites

• Python 3.10+
• uv package manager
• AWS Account with Bedrock access
• Claude Desktop app

Setup

1. Clone the repository:

git clone https://github.com/yourusername/recursive-companion-mcp.git
cd recursive-companion-mcp

2. Install dependencies:

uv sync

3. Configure AWS credentials as environment variables or through AWS CLI

4. Add to Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "recursive-companion": {
      "command": "/path/to/recursive-companion-mcp/run_server.sh",
      "env": {
        "AWS_REGION": "us-east-1",
        "AWS_ACCESS_KEY_ID": "your-key",
        "AWS_SECRET_ACCESS_KEY": "your-secret",
        "BEDROCK_MODEL_ID": "anthropic.claude-3-sonnet-20240229-v1:0",
        "CRITIQUE_MODEL_ID": "anthropic.claude-3-haiku-20240307-v1:0",
        "CONVERGENCE_THRESHOLD": "0.95",
        "PARALLEL_CRITIQUES": "2",
        "MAX_ITERATIONS": "5",
        "REQUEST_TIMEOUT": "600"
      }
    }
  }
}

Usage

The tool provides several MCP endpoints:

Start a refinement session

Use start_refinement to refine: "Explain the key principles of secure API design"

Continue refinement step by step

Use continue_refinement with session_id "abc123..."

Get final result

Use get_final_result with session_id "abc123..."

Other tools

• get_refinement_status - Check progress without advancing
• list_refinement_sessions - See all active sessions

Configuration

Environment Variable	Default	Description
BEDROCK_MODEL_ID	anthropic.claude-3-sonnet-20240229-v1:0	Main generation model
CRITIQUE_MODEL_ID	Same as BEDROCK_MODEL_ID	Model for critiques (use Haiku for speed)
CONVERGENCE_THRESHOLD	0.98	Similarity threshold for convergence (0.90-0.99)
PARALLEL_CRITIQUES	3	Number of parallel critiques per iteration
MAX_ITERATIONS	10	Maximum refinement iterations
REQUEST_TIMEOUT	300	Timeout in seconds

Performance

With optimized settings:

• Each iteration: 60-90 seconds
• Typical convergence: 2-3 iterations
• Total time: 2-4 minutes (distributed across multiple calls)

Using Haiku for critiques reduces iteration time by ~50%.

Architecture

┌─────────────┐     ┌──────────────┐     ┌─────────────┐
│   Claude    │────▶│  MCP Server  │────▶│   Bedrock   │
│  Desktop    │◀────│              │◀────│   Claude    │
└─────────────┘     └──────────────┘     └─────────────┘
                           │
                           ▼
                    ┌──────────────┐
                    │   Session    │
                    │   Manager    │
                    └──────────────┘

Development

Running tests

uv run pytest tests/

Local testing

uv run python test_incremental.py

Attribution

This project is inspired by recursive-companion by Hank Besser. The original implementation provided the conceptual Draft → Critique → Revise → Converge pattern. This MCP version adds:

• Session-based incremental processing to avoid timeouts
• AWS Bedrock integration for Claude and Titan embeddings
• Domain auto-detection and specialized prompts
• Mathematical convergence measurement
• Support for different models for critiques vs generation

Contributing

Contributions are welcome! Please read our Contributing Guide for details.

License

MIT License - see LICENSE file for details.

Acknowledgments

• Original concept: Hank Besser's recursive-companion
• Built for the Model Context Protocol
• Uses AWS Bedrock for LLM access
• Inspired by iterative refinement patterns in AI reasoning