MCP Reasoner

🚀 MCP Reasoner

An inference implementation designed for Claude Desktop, enabling it to use both beam search and Monte Carlo Tree Search (MCTS) simultaneously. Honestly, this project initially aimed to see if we could make Claude perform better in complex problem-solving... and it turns out we can.

Current Version:

v2.0.0

New Features:

Two experimental inference algorithms have been added:

- `mcts-002-alpha`

    - Uses the A* search method and combines it with the policy simulation layer from the early *alpha* implementation.

    - Also includes the adaptive exploration simulator and the result-based inference simulator from the early *alpha* implementation.

*Note*: The implementation of these simulators is not yet complete and may change.

- `mcts-002alt-alpha`

    - Uses the bidirectional search method and combines it with the policy simulation layer from the early *alpha* implementation.

    - Also includes the adaptive exploration simulator and the result-based inference simulator from the early *alpha* implementation.

*Note*: The implementation of these simulators is not yet complete and may change.

Why were mcts-001-alpha and mcts-001alt-alpha abandoned?

In short: They were almost useless, being nearly similar to the basic mcts method. After initial testing, the results showed that the basic thinking processes had similar effects, indicating that simply adding policy simulation might not make a difference.

Then why introduce the policy simulation layer now?

Well, I think combining strategy and search is the key, as most algorithms are implemented in this way.

Previous Version:

v1.1.0

Added model control over search parameters:

beamWidth - Allows Claude to adjust the number of different paths to track (1 - 10).

numSimulations - Fine-tunes the number of MCTS simulations (1 - 150).

✨ Features

• Two switchable search strategies:
  • Beam search (suitable for simple problems).
  • MCTS (when problems become complex) and its alpha variants (see above).
• Track the quality of different inference paths.
• Visualize different ways Claude solves problems.
• Analyze the performance of the inference process.
• Comply with the MCP protocol (obviously).

📦 Installation

git clone https://github.com/frgmt0/mcp-reasoner.git 

Or clone the original repository:

git clone https://github.com/Jacck/mcp-reasoner.git

cd mcp-reasoner
npm install
npm run build

🛠️ Configuration

Add to the Claude Desktop configuration:

{
  "mcpServers": {
    "mcp-reasoner": {
      "command": "node",
      "args": ["path/to/mcp-reasoner/dist/index.js"],
    }
  }
}

🧪 Testing

[More tests are coming soon]

⏱️ Benchmarking

[Benchmarks will be added soon]

Key benchmarks to be tested:

• MATH500
• GPQA-Diamond
• GMSK8
• Possibly Polyglot and/or SWE-Bench

📄 License

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