Structured ThinkingSAFE

A TypeScript Model Context Protocol (MCP) server based on Arben Ademi's Sequential Thinking Python server. The motivation for this project is to allow LLMs to programmatically construct mind maps to explore an idea space, with enforced "metacognitive" self-reflection.

Setup

Set the tool configuration in Claude Desktop, Cursor, or another MCP client as follows:

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

Overview

Thought Quality Scores

When an LLM captures a thought, it assigns that thought a quality score between 0 and 1. This score is used, in combination with the thought's stage, for providing "metacognitive" feedback to the LLM how to "steer" its thinking process.

Thought Stages

Each thought is tagged with a stage (e.g., Problem Definition, Analysis, Ideation) to help manage the life-cycle of the LLM's thinking process. In the current implementation, these stages play a very important role. In effect, if the LLM spends too long in a given stage or is having low-quality thoughts in the current stage, the server will provide feedback to the LLM to "steer" its thinking toward other stages, or at least toward thinking strategies that are atypical of the current stage. (E.g., in deductive mode, the LLM will be encouraged to consider more creative thoughts.)

Thought Branching

The LLM can spawn “branches” off a particular thought to explore different lines of reasoning in parallel. Each branch is tracked separately, letting you manage scenarios where multiple solutions or ideas should coexist.

Memory Management

The server maintains a "short-term" memory buffer of the LLM's ten most recent thoughts, and a "long-term" memory of thoughts that can be retrieved based on their tags for summarization of the entire history of the LLM's thinking process on a g