Reasoning Engine MCP Server

MCP (Model Context Protocol) server for deep algorithmic and mathematical reasoning via Hermes Agent.

Overview

reasoning_engine is a generic OpenAI-compatible chat completions wrapper that calls any LLM endpoint to produce only the final distilled answer — no chain-of-thought, no reasoning traces exposed. Designed for complex computational problems where you need the result, not the process.

By default it's configured for vibethinker-3b via LM Studio, but can target any model on any endpoint — per-call or server-level.

Related MCP server: verifiable-thinking-mcp

Use Cases

• Algorithm design & optimization

• Data structure complexity analysis

• Competitive programming challenges

• Mathematical proofs & derivations

• Multi-step logical deduction tasks

• Any deep technical reasoning requiring compressed output

Installation

Prerequisites

• Python 3.10+

• LM Studio running with vibethinker-3b model loaded (or any OpenAI-compatible LLM server)

Setup

# Test the server directly
python3 src/reasoning_engine_mcp/server.py --print-config

Configuration

Add to your Hermes Agent ~/.hermes/config.yaml under mcp_servers:

mcp_servers:
  reasoning_engine:
    command: python3
    args:
    - /path/to/reasoning-engine-mcp/src/reasoning_engine_mcp/server.py
    env:
      REASONING_ENGINE_MODEL: vibethinker-3b
      REASONING_ENGINE_BASE_URL: http://127.0.0.1:1234/v1

Server-Level Environment Variables

Per-Call Overrides

The tool accepts additional parameters that override the server-level defaults for a single call:

Resolution order: tool params → env vars → hardcoded defaults.

Usage

Once configured, the reasoning_engine tool is available in Hermes Agent sessions. Call it with a detailed problem prompt:

# Example: Algorithm design (using server-level defaults)
call_tool("reasoning_engine", {
    "prompt": """
    Problem: Find the minimum number of operations to convert array A to array B
    where each operation can increment or decrement an element by 1.
    
    Constraints:
    - Array length up to 10^5
    - Elements range from -10^9 to 10^9
    
    Input Format: Two arrays of equal length
    Output Format: Single integer (minimum operations)
    """,
    "max_tokens": 4096,
    "temperature": 0.3
})

# Example: Override model per-call
call_tool("reasoning_engine", {
    "prompt": "Prove that the sum of two even numbers is always even.",
    "model": "qwen/qwen3.6-35b-a3b",
    "max_tokens": 4096,
    "temperature": 0.1
})

# Example: Use OpenAI API instead of local server
call_tool("reasoning_engine", {
    "prompt": "Design an O(n log n) sorting algorithm with detailed complexity analysis.",
    "model": "anthropic/claude-sonnet-4",
    "base_url": "https://api.openai.com/v1",
    "api_key": "sk-...",
    "max_tokens": 8192,
    "temperature": 0.5
})

Core Parameters

Architecture

Hermes Agent
    │
    ▼
MCP Client (stdio transport)
    │
    ▼
reasoning-engine-mcp server.py
    │  (reads JSON-RPC from stdin)
    │  (writes JSON-RPC to stdout)
    │
    ▼
OpenAI-compatible API (HTTP POST /chat/completions)
    │
    ▼
Any LLM model (vibethinker-3b, Claude, GPT, Qwen, etc.)

The MCP server is a lightweight stdio bridge — no external dependencies, pure Python standard library.

Language Enforcement

All responses are strictly in English regardless of the prompt language. This is enforced via:

1. System prompt with explicit LANGUAGE ENFORCEMENT section

2. Footer appended to every user message

License

MIT License — feel free to fork, modify, and share.