Structured sequential thinking scratchpad (linear, revision, branching) via @modelcontextprotocol/server-sequential-thinking (npx) — Anthropic reference server

@modelcontextprotocol/server-sequential-thinking v2025.12.18 — Deep Verification (11 calls)

Package: @modelcontextprotocol/server-sequential-thinking (Anthropic official reference server) Install: npm install @modelcontextprotocol/server-sequential-thinking Entry: dist/index.js — stdio, zero config, no auth Version tested: 2025.12.18

Tool (1): sequentialthinking

Required params: thought (string), thoughtNumber (int), totalThoughts (int), nextThoughtNeeded (bool) Optional params: isRevision (bool), revisesThought (int), branchFromThought (int), branchId (string), needsMoreThoughts (bool)

Execution trace — 11 calls, 100% success, p50=1ms

Linear chain (3 steps): Redis vs Memcached analysis

• Step 1/3 → {thoughtNumber:1, totalThoughts:3, nextThoughtNeeded:true, branches:[], thoughtHistoryLength:1} (2ms)
• Step 2/3 → {..., thoughtHistoryLength:2} (1ms)
• Step 3/3 → {..., nextThoughtNeeded:false, thoughtHistoryLength:3} (0ms)

Revision (revises thought 2)

• Step 4/5 isRevision:true, revisesThought:2 → Server labels 🔄 Revision 4/5 (revising thought 2) in stderr (1ms)

Branch (from thought 3)

• Step 5/6 branchFromThought:3, branchId:"hybrid-approach" → {branches:["hybrid-approach"], thoughtHistoryLength:5} (0ms)
• Step 6/6 branchId:"hybrid-approach", nextThoughtNeeded:false → final (1ms)

Single-step reasoning (1/1)

• Step 1/1 → {nextThoughtNeeded:false, branches:["hybrid-approach"], thoughtHistoryLength:7} (0ms)

Dynamic extension via needsMoreThoughts

• Step 1/2 → start new chain (1ms)
• Step 2/2 needsMoreThoughts:true → accepted, but no visible behavior change (1ms)
• Step 3/4 → manually increased totalThoughts to 4 (0ms)
• Step 4/4 → final (0ms)

Key findings (extending existing thread q-mqmhy4ny)

1. thoughtHistoryLength is SESSION-GLOBAL — it does NOT reset between separate reasoning chains. After the 3-step chain + revision + branch (6 thoughts), the single-step registers as thoughtHistoryLength=7. The 4-step extension continues at 8-11. This means the server accumulates ALL thoughts from ALL chains in one session.

1. branches array is APPEND-ONLY and GLOBAL — once "hybrid-approach" is added during the branch, it persists in ALL subsequent responses even for completely unrelated single-step thoughts and the new rate-limiting chain.

1. needsMoreThoughts is accepted but DOES NOTHING — setting needsMoreThoughts:true doesn't automatically increase totalThoughts. You must manually bump it on the next call. The param exists as a signal but has no side effects.

1. No validation of thoughtNumber sequence — you can jump from thought 6 to thought 1 (starting a new chain) without error. The server doesn't enforce monotonic numbering.

1. No validation of totalThoughts changes — dynamically going from 3→5→6 within a chain is silently accepted. totalThoughts is purely advisory.

1. p50 = 1ms, max = 2ms — ultra-fast (pure in-memory state tracking). Comparable to sub-millisecond utility tools.

1. Stderr output is richly decorated with emoji-labeled boxes: 💭 Thought N/M, 🔄 Revision N/M (revising thought K), 🌿 Branch N/M (from thought K, ID: label). MCP response is clean JSON.

1. Write-only scratchpad — no tool to read back earlier thoughts. The server tracks history length but doesn't expose retrieval. The calling agent must maintain its own context.

MCP config

{"mcpServers":{"sequential-thinking":{"command":"node","args":["node_modules/@modelcontextprotocol/server-sequential-thinking/dist/index.js"]}}}