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

@modelcontextprotocol/server-sequential-thinking v2025.12.18 — Verified Recipe

Package: @modelcontextprotocol/server-sequential-thinking (Anthropic official reference server) Install: npm install @modelcontextprotocol/server-sequential-thinking Entry: dist/index.js — runs on stdio, stateful within a server session (no file persistence) Version tested: 2025.12.18

Tools (1)

Parameters detail

• `thought` (string, required): The content of this thinking step
• `thoughtNumber` (int, required): Current step number (1-indexed)
• `totalThoughts` (int, required): Expected total steps — can increase dynamically
• `nextThoughtNeeded` (bool, required): true to continue chain, false to end
• `isRevision` (bool): Mark this thought as revising a prior one
• `revisesThought` (int): Which thought number is being revised
• `branchFromThought` (int): Fork from a specific earlier thought
• `branchId` (string): Label for the branch
• `needsMoreThoughts` (bool): Signal that totalThoughts should increase

Execution trace — 6 calls, 100% success

Tested a multi-step caching architecture design with revision and branching:

1. Thought 1/5 — Define problem (API caching at 10K req/s)

→ Response: {thoughtNumber:1, totalThoughts:5, nextThoughtNeeded:true, branches:[], thoughtHistoryLength:1} → Server prints 💭 Thought 1/5 box to stderr

1. Thought 2/5 — Analyze cache-aside vs stale-while-revalidate

→ {thoughtNumber:2, totalThoughts:5, branches:[], thoughtHistoryLength:2}

1. Revision 3/5 (revising thought 2) — isRevision:true, revisesThought:2

→ Server prints 🔄 Revision 3/5 (revising thought 2) to stderr → {thoughtNumber:3, totalThoughts:5, branches:[], thoughtHistoryLength:3}

1. Branch 4/6 (from thought 2, ID: "cdn-approach") — branchFromThought:2, branchId:"cdn-approach", totalThoughts:6

→ Server prints 🌿 Branch 4/6 (from thought 2, ID: cdn-approach) to stderr → {thoughtNumber:4, totalThoughts:6, branches:["cdn-approach"], thoughtHistoryLength:4} → Note: totalThoughts dynamically increased from 5→6

1. Thought 5/6 — Synthesize two-layer approach

→ {thoughtNumber:5, totalThoughts:6, branches:["cdn-approach"], thoughtHistoryLength:5}

1. Thought 6/6 — Final decision, nextThoughtNeeded:false

→ {thoughtNumber:6, totalThoughts:6, nextThoughtNeeded:false, branches:["cdn-approach"], thoughtHistoryLength:6}

Key observations

1. Single tool, rich semantics — all reasoning operations (think, revise, branch, extend) use the same sequentialthinking tool with different flag combinations.
2. Dynamic chain length — totalThoughts can increase mid-chain (went from 5→6 when branching).
3. Revision tracking — isRevision:true + revisesThought:N marks a thought as correcting a prior one. Server labels it 🔄 Revision in stderr output.
4. Branching — branchFromThought:N + branchId:"label" forks the reasoning. Branch IDs accumulate in the response's branches array.
5. Stderr decoration — server prints formatted boxes with emoji (💭, 🔄, 🌿) to stderr for human-readable logging. MCP response is clean JSON.
6. No persistence — thought chain lives in server memory only. Restarting the server loses all thoughts.
7. No retrieval tool — there is no way to read back earlier thoughts via MCP. The server tracks thoughtHistoryLength but doesn't expose a getThoughts tool. The agent must maintain its own context.
8. Response shape — always returns {thoughtNumber, totalThoughts, nextThoughtNeeded, branches[], thoughtHistoryLength}.

Key gotchas

1. **⚠️ No re