Cotforce Mcp

Name: Cotforce Mcp
Author: islobodan

@islobodan

关于 Cotforce Mcp

Give brains to your small models. MCP server that enforces step-by-step Chain-of-Thought — turns 4B models into methodical reasoners

#mcp-cot-chain-of-thought

基本信息

分类

其他

传输方式

stdio

发布者

islobodan

提交者

Slobodan Ivković

配置

使用下面的配置,将此服务器添加到你的 MCP 客户端。

{
  "mcpServers": {
    "cotforce": {
      "command": "npx",
      "args": [
        "-y",
        "@slbdn/cotforce-mcp"
      ],
      "env": {
        "MODE": "direct",
        "API_BASE_URL": "http://localhost:1234/v1",
        "MODEL": "local-model",
        "MAX_RETRIES": "2"
      }
    }
  }
}

工具

未检测到工具

工具是从 README 中自动提取的。维护者可以在 ## Tools 标题下列出工具,即可填充这部分内容。

概览

What is Cotforce Mcp?

Cotforce Mcp is a Model Context Protocol server that enforces step-by-step Chain-of-Thought (CoT) reasoning on any LLM call. It integrates with MCP clients (Claude Desktop, Cursor) via sampling or with any OpenAI‑compatible API via direct HTTP. It is designed for users who want to improve reasoning in small or large language models without fine‑tuning.

How to use Cotforce Mcp?

Install the npm package @slbdn/cotforce-mcp (requires Node.js ≥18). Configure via environment variables in your MCP client settings. Add a server entry like "cotforce": { "command": "npx", "args": ["-y", "@slbdn/cotforce-mcp"], "env": { "MODEL": "claude-3-5-sonnet" } } to claude_desktop_config.json. Then call the solve_problem tool with a prompt argument.

Key features of Cotforce Mcp

Rigid CoT enforcement with valid JSON output
Adaptive multi‑layer parser pipeline (5 built‑in parsers)
Zod runtime validation for tool arguments and CoT results
Automatic retry with increasing temperature (up to 3 attempts)
Token budgeting with tiktoken and configurable overhead
Model‑specific system prompts for Claude, GPT‑4, Gemini, Grok
Structured metrics and comprehensive test suite (151 tests)

Use cases of Cotforce Mcp

Solving complex cryptarithmetic puzzles like SEND + MORE = MONEY with small models
Improving reasoning on math, logic, and word problems
Performing code analysis, regex generation, and SQL queries with structured output
Validating results against a user‑supplied schema with automatic retry

FAQ from Cotforce Mcp

What LLMs does Cotforce Mcp work with?

It works with any LLM accessible via MCP sampling or direct HTTP, including Claude, GPT‑4, Gemini, Grok, and local models via LMStudio or Ollama.

What are the runtime requirements?

Cotforce Mcp requires Node.js ≥18. For direct HTTP mode, no additional services are needed if you provide an API key or use a local endpoint.

How do I handle truncated responses mid‑reasoning?

Increase REASONING_OVERHEAD (default 800) or reduce parsers via COT_PARSERS to save token budget for reasoning.

Why does my MCP client time out before getting a result?

Complex CoT reasoning can take 60–90 seconds. Increase Cotforce’s TIMEOUT variable and also check your MCP client’s own timeout setting (e.g., LM Studio, Cursor).

How does Cotforce Mcp enforce structured output?

It uses strict system prompts with few‑shot examples, a priority‑sorted parser pipeline, and Zod schemas to extract and validate {reasoning, result} JSON from any LLM response.