Structured sequential reasoning with revision and branching via @modelcontextprotocol/server-sequential-thinking (npx) — Anthropic reference server

@modelcontextprotocol/server-sequential-thinking v2025.12.18 — verified recipe

Install & spawn:

npm install --prefix /tmp/seq-thinking @modelcontextprotocol/server-sequential-thinking

Entry point: node_modules/@modelcontextprotocol/server-sequential-thinking/dist/index.js

Tool: sequentialthinking

Required params: thought (string), nextThoughtNeeded (boolean), thoughtNumber (integer ≥1), totalThoughts (integer ≥1)

Optional params: isRevision (bool), revisesThought (int), branchFromThought (int), branchId (string), needsMoreThoughts (bool)

Returns: {thoughtNumber, totalThoughts, nextThoughtNeeded, branches: string[], thoughtHistoryLength: number}

Verified behaviors (10 calls, 100% success, p50=1ms):

1. Linear chain: steps 1→2→3 with nextThoughtNeeded: true/true/false — history increments 1→2→3
2. Revision: step 4 with isRevision: true, revisesThought: 3 — server accepts, totalThoughts can be adjusted upward simultaneously
3. Branching: step 5 with branchFromThought: 2, branchId: "reverse-sorted" — branch appears in branches[] array in all subsequent responses
4. Dynamic totalThoughts: changed from 3→5→6 mid-chain — server accepts without error
5. New problem (thoughtNumber reset to 1): history does NOT reset — thoughtHistoryLength keeps incrementing (7 after step 1 of new problem). Branches from prior problems persist too.
6. Minimal call: {thought: "Simple test.", nextThoughtNeeded: false, thoughtNumber: 1, totalThoughts: 1} — works
7. Long text: 4500+ chars in thought — accepted without truncation

Key gotchas:

• Server is a PURE STATE TRACKER — it does NOT reason, validate, or transform thought content. It stores thoughts and returns metadata (number, branches, history length). The LLM does the actual thinking.
• History is GLOBAL and PERSISTENT across all calls in a session — thoughtHistoryLength never resets. Starting a "new problem" with thoughtNumber: 1 doesn't clear prior state.
• Branches persist forever — once branchId: "foo" is created, it appears in every subsequent response's branches[]
• No `reset` or `clear` tool — the only way to start fresh is to restart the server process
• Server prints formatted boxes to stderr (visible in terminal, not in MCP response) — cosmetic only
• `needsMoreThoughts: true` is accepted but has no observable effect on the response (the caller controls totalThoughts directly)
• No maximum thought limit enforced by the server (schema says maximum: 9007199254740991)
• Anthropic reference server — canonical implementation for Claude Desktop, used as structured scratchpad for chain-of-thought