think-mcp-server

Name: think-mcp-server
Author: marcopesani

@marcopesani

リポジトリ

think-mcp-server について

A minimal MCP Server based on the Anthropic's "think" tool research

基本情報

カテゴリ

その他

ライセンス

MIT

ランタイム

node

トランスポート

stdio

公開者

marcopesani

設定

以下の設定を使って、このサーバーを MCP 対応クライアントに追加してください。

{
  "mcpServers": {
    "think-mcp-server-marcopesani": {
      "command": "docker",
      "args": [
        "build",
        "-t",
        "think-mcp-server",
        "."
      ]
    }
  }
}

ツール

Enables Claude to think about a problem or analyze information

概要

What is think-mcp-server?

think-mcp-server is a minimal MCP (Model Context Protocol) server that provides Claude AI models with the “think” tool capability, based on Anthropic’s research published on March 20, 2025. It enables Claude to pause during response generation for an additional reasoning step, improving performance on complex, multi-step tool‑use tasks.

How to use think-mcp-server?

Install via npm (npm install -g think-mcp-server) or Smithery, then add the server configuration to your AI client (Claude Desktop, Cline, Cursor, or Docker). In your prompts, include instructions for when and how Claude should use the “think” tool. For best results, place detailed guidance in the system prompt rather than the tool description.

Key features of think-mcp-server

Implements the “think” tool for Claude’s reasoning steps
Designed for complex, multi-step tool‑use tasks
Communicates over stdio using the MCP protocol
Logs thinking steps without changing the external environment
Provides a structured input schema requiring only a “thought” string
Backed by Anthropic’s published research showing significant performance gains

Use cases of think-mcp-server

Analyzing tool output before making the next tool call
Following detailed guidelines in policy‑heavy environments
Making careful, sequential decisions where mistakes are costly
Adding explicit reasoning steps in TypeScript development workflows

FAQ from think-mcp-server

What exactly does the “think” tool do?

The “think” tool gives Claude a dedicated space to pause during response generation and include an additional thinking step. It does not fetch new information or change any database; it simply appends the thought to a log.

How is it different from extended thinking?

Extended thinking happens before response generation, whereas the “think” tool allows Claude to pause during response generation to consider whether it has all necessary information to proceed.

What are the runtime requirements and dependencies?

think-mcp-server is written in TypeScript and runs as a Node.js MCP server. Dependencies are managed via npm. It uses stdio for communication, so a Node.js runtime is required.

How do I debug the server?

Debugging is best done using the MCP Inspector: run npm run inspector and open the provided URL in a browser.

What performance improvements have been measured?

Anthropic’s evaluations showed a 54% relative improvement on τ-Bench airline domain (pass^1: 0.570 vs 0.370 baseline) and a 1.6% average improvement on SWE-bench, along with enhanced consistency across multiple trials.