mcp-sequentialthinking-tools

To use this MCP server with Claude Desktop or other compatible MCP clients, add the following configuration to your claude_desktop_config.json. The server is executed via npx and supports an optional environment variable to control the maximum history size per session.

Environment Variables:

• MAX_HISTORY_SIZE: The maximum number of records stored per session. Defaults to 1000.

{
	"mcpServers": {
		"mcp-sequentialthinking-tools": {
			"command": "npx",
			"args": ["-y", "mcp-sequentialthinking-tools"],
			"env": {
				"MAX_HISTORY_SIZE": "1000"
			}
		}
	}
}

Use clear_thinking_history to wipe reasoning records.

Parameters:

• session_id: The ID of the session to clear (defaults to default).
• all_sessions: If set, clears every session history bucket.

Use get_thinking_history to retrieve stored thoughts for a specific session. This is useful for inspecting the reasoning path during long agent runs.

Parameters:

• session_id: The ID of the session to retrieve (defaults to default).
• branch_id: (Optional) A filter to return only thoughts from a specific branch.
• limit: The maximum number of records to return. Default is 50, maximum is 500.

You can control the capacity of the thinking store by setting the MAX_HISTORY_SIZE environment variable.

• Default: 1000
• Behavior: If the environment variable is not a finite positive integer, it falls back to the default value.

The sequential-thinking-guidance prompt provides instructions to the model on how to use the sequential thinking tools correctly. It emphasizes:

• Using the tools only for problems benefiting from explicit multi-step reasoning.
• Keeping thoughts short and revising/branching when evidence changes.
• Passing available_tools and recommended_tools for validation.
• Not claiming the server performs the reasoning; the model is the author of the plan.

When interacting with the sequential thinking tools, tools can be referenced either as a simple string (the tool name) or as an object containing a name and an optional description.

Tool recommendations provide structured guidance for the model, including the tool_name, a confidence score, a rationale for the suggestion, a priority level, suggested_inputs (a record of key-value pairs), and a list of alternatives.

// Example of a tool_reference object
const toolRef: tool_reference = {
  name: "fetch_url",
  description: "Retrieves content from a web page"
};

// Example of a tool_recommendation
const recommendation: tool_recommendation = {
  tool_name: "search_web",
  confidence: 0.95,
  rationale: "The user is asking about current events.",
  priority: 1,
  suggested_inputs: { query: "latest news" },
  alternatives: ["google_search", "bing_search"]
};

The thinking_store class manages sequential thinking sessions. You can initialize it with an optional thinking_store_options object to control the maximum number of thoughts retained per session.

If max_history_size is not provided, it defaults to 1000. The value is sanitized to ensure it is at least 1.

import { thinking_store } from './thinking.js';

const store = new thinking_store({
  max_history_size: 500
});

The sequential thinking process relies on thought_input and thought_record objects.

thought_input is used to submit a new step in the reasoning process. It includes the thought text, the current thought_number, the total_thoughts expected, and flags like next_thought_needed or needs_more_thoughts. It also supports branching logic via revises_thought, branch_from_thought, and branch_id. You can also provide available_tools and recommended_tools to guide the model.

thought_record is the persisted version of a thought, which adds a session_id and a created_at timestamp.

const input: thought_input = {
  thought: "I need to check the weather in London.",
  thought_number: 1,
  total_thoughts: 5,
  next_thought_needed: true,
  available_tools: ["get_weather"],
  remaining_steps: ["Check temperature", "Check humidity"]
};