MCP Structured Thinking

🚀 Structured Thinking MCP Server

A TypeScript-based Model Context Protocol (MCP) server inspired by Arben Ademi's Sequential Thinking Python server. This project aims to allow Large Language Models (LLMs) to programmatically construct mind maps to explore an idea space and guide their thinking process by enforcing "metacognitive" self-reflection.

🚀 Quick Start

Set up the tool configuration in Claude Desktop, Cursor, or other MCP clients as follows:

{
  "structured-thinking": {
    "command": "npx",
    "args": ["-y", "structured-thinking"]
  }
}

📚 Documentation

Thought Quality Score

When an LLM captures a thought, it assigns a quality score to the thought, ranging from 0 to 1. This score is combined with the current phase to provide "metacognitive" feedback to the LLM, guiding how its thinking process should "turn".

Thought Phase

Each thought is labeled with a phase (e.g., problem definition, analysis, ideation) to help manage the lifecycle of the LLM's thinking process. In the current implementation, these phases play a crucial role. Essentially, if the LLM stays in a certain phase for too long or generates low-quality thoughts in the current phase, the server provides feedback to the LLM, guiding its thinking to shift to other phases or at least encouraging it to adopt thinking strategies that are less typical of the current phase. (For example, in the deductive mode, the LLM will be encouraged to consider more creative thinking.)

Thought Branching

An LLM can "spawn" "branches" from a specific thought to explore different lines of reasoning in parallel. Each branch is tracked separately so that you can manage multiple solutions or ideas that need to coexist.

Memory Management

The server maintains a "short-term" memory buffer that records the LLM's ten most recent thoughts and a "long-term" memory library that can retrieve thoughts based on tags, used to summarize the history of the LLM's thinking process on the entire topic.

🔧 Technical Details

Simple Implementation of Metacognitive Monitoring

Currently, the quality metric and metacognitive feedback are mechanically derived from applying a phase-based multiplier to a single self-reported quality score.

As part of future work, I plan to add more sophisticated metacognitive feedback, including semantic analysis of thought content, a thought validation process, and smarter monitoring of reasoning errors.

Lack of User Interface

Currently, the server stores all thoughts in memory and does not persist them to a file or database. There is also no user interface for reviewing the idea space or visualizing the mind map.

As part of future work, I plan to integrate a simple visualization client so that users can watch the development of the thought map.

📦 API Documentation

The server exposes the following MCP tools:

capture_thought

Create a thought in the thought history and record the thought's type, quality, content, and its relationship with other thoughts.

Parameters:

• thought: The content of the current thought
• thought_number: The current position in the sequence
• total_thoughts: The expected total number of thoughts
• next_thought_needed: Whether another thought is needed
• current_phase: The current phase (e.g., "Define Problem", "Analysis", "Ideation")
• related_tags: A list of related tags

revise_thought

Revise an existing thought.

Parameters:

• thought_id: The ID of the thought to be revised
• new_content: The new content
• new_quality_score: The new quality score (optional)

retrieve_relevant_thoughts

Retrieve relevant thoughts based on the current context.

Parameters:

• current_context: The current context
• max_results: The maximum number of results

summarize_thought_process

Get a summary of the entire thinking process.

Parameters:

• start_time: The start time (optional)
• end_time: The end time (optional)

💻 Usage Examples

Basic Usage

Problem Definition Phase

{
  "action": "capture_thought",
  "thought": "We need to solve the user's problem. First, clarify what the problem is.",
  "thought_number": 1,
  "total_thoughts": 5,
  "next_thought_needed": true,
  "current_phase": "Define Problem",
  "related_tags": ["Problem Definition"]
}

Analysis Phase

{
  "action": "capture_thought",
  // Here should be the content for analysis phase in the original text, but it's incomplete in your input.
}