Sequential Thinking

mcp-sequentialthinking-tools

A lightweight MCP server for recording sequential reasoning steps. It is a scratchpad with history, branching, revision metadata, and optional validation for model-authored tool plans.

It does not discover your other MCP tools and it does not choose tools for the model. If you pass available_tools and recommended_tools, the server validates that the recommended names exist and stores the step.

Why use it?

Use this when a task benefits from explicit, inspectable reasoning:

• breaking a messy problem into steps;
• revising or branching a plan;
• keeping a small reasoning history by session;
• validating tool-plan names against a supplied tool list;
• clearing or inspecting reasoning history during a long agent run.

Do not use it for trivial requests. It adds overhead.

Tools

sequentialthinking_tools

Records one thought.

Required parameters:

• thought — current reasoning step
• thought_number — current step number
• total_thoughts — current estimate; automatically raised if lower than thought_number
• next_thought_needed — whether another thought is needed

Optional parameters:

• session_id — history bucket; defaults to default
• is_revision, revises_thought
• branch_from_thought, branch_id
• needs_more_thoughts
• available_tools — array of tool names or { name, description } objects
• recommended_tools — model-authored recommendations to validate/store
• remaining_steps — short list of upcoming steps

Example:

{
	"session_id": "svelte-debug",
	"thought": "First inspect the route files, then run the failing check.",
	"thought_number": 1,
	"total_thoughts": 3,
	"next_thought_needed": true,
	"available_tools": ["read", "bash"],
	"recommended_tools": [
		{
			"tool_name": "read",
			"confidence": 0.9,
			"rationale": "Need to inspect the relevant files before editing.",
			"priority": 1
		}
	]
}

If recommended_tools contains a name not present in available_tools, the call returns isError: true and does not store the thought.

Security posture

The server treats thought text, tool descriptions, rationales, and remaining-step text as untrusted input. Prompt-injection-like text is scanned and redacted before it is stored or returned in history. Calls with redactions include security_warnings showing which fields matched.

This is defensive filtering, not a guarantee that arbitrary adversarial text is safe. Do not put secrets in thoughts or tool descriptions.

get_thinking_history

Returns stored thoughts for a session.

Parameters:

• session_id — defaults to default
• branch_id — optional branch filter
• limit — max records to return; default 50, max 500

clear_thinking_history

Clears one session or every session.

Parameters:

• session_id — defaults to default
• all_sessions — clear all history buckets

Prompt

sequential-thinking-guidance

A short prompt that tells the model how to use this server honestly: as a scratchpad and validator, not as an external reasoning engine.

Configuration

Claude Desktop / compatible MCP clients

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

MAX_HISTORY_SIZE is per session and defaults to 1000.

The server uses tmcp and includes a small stdio transport that accepts both standard Content-Length framed MCP messages and newline-delimited JSON used by older tmcp tooling.

Development

pnpm install
pnpm test
pnpm build
pnpm check

This project uses vite-plus for build, test, format, and lint orchestration.

Publishing

pnpm changeset
pnpm changeset version
pnpm release

License

MIT License — see LICENSE.

Acknowledgments

• Built on the Model Context Protocol
• Adapted from the MCP Sequential Thinking Server