mcp-sequentialthinking-tools

Guides problem-solving by breaking down complex problems and recommending the best MCP tools for each step.

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 doesnotdiscover your other MCP tools and it doesnotchoose tools for the model. If you passavailable_toolsandrecommended_tools, the server validates that the recommended names exist and stores the step.

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.

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

- session_id— history bucket; defaults todefault
- 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

{ "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 } ] }

Ifrecommended_toolscontains a name not present inavailable_tools, the call returnsisError: trueand does not store the thought.

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 includesecurity_warningsshowing 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.

- session_id— defaults todefault
- branch_id— optional branch filter
- limit— max records to return; default50, max500

- session_id— defaults todefault
- all_sessions— clear all history buckets

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

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

MAX_HISTORY_SIZEis per session and defaults to1000.

The server usestmcpand includes a small stdio transport that accepts both standardContent-Lengthframed MCP messages and newline-delimited JSON used by oldertmcptooling.

pnpm install pnpm test pnpm build pnpm check

This project usesvite-plusfor build, test, format, and lint orchestration.

pnpm changeset pnpm changeset version pnpm release

- Built on theModel Context Protocol
- Adapted from theMCP Sequential Thinking Server

Sync Calendars, Scheduling Links, AI Executive Scheduling Assistant, Unified Calendar