What it is
Sequential Thinking is one of the seven reference servers still maintained in modelcontextprotocol/servers; it has not been moved to servers-archived. It exposes one tool, registered as sequentialthinking (the README calls it sequential_thinking, a mismatch tracked in issue #5204). The model calls it repeatedly, each time recording one numbered thought with an estimate of how many remain, and can mark a thought as revising an earlier one or branch into an alternative line.
The server computes nothing. It keeps the thought history in memory for the length of the connection and logs each thought to stderr unless DISABLE_THOUGHT_LOGGING is set. Run it with npx -y @modelcontextprotocol/server-sequential-thinking, which needs Node.js 24 or later since version 2.0.0 (19 October 2026), or the mcp/sequentialthinking Docker image. No key is involved.
Install
npx -y @modelcontextprotocol/server-sequential-thinkingOur review
On a current reasoning model this server mostly adds cost. Issue #4507 measures its single tool definition at about 921 tokens, paid in every session whether or not the tool is called, and each thought is a full tool round trip. Models from Anthropic, OpenAI and Google with extended thinking already plan and revise before they answer; routing that through tool calls duplicates work done internally.
It still has two honest uses. With smaller or non-reasoning models, the schema's fields (thoughtNumber, totalThoughts, isRevision, branchFromThought) push the model to break a problem down and revisit steps it would otherwise skip. And because every thought lands in the tool log, it leaves a reviewable trace of how an agent reached a decision, which a model's internal reasoning often does not.
Nothing in the docs claims better answers. The README's test for whether it is working is simply that the host calls the tool several times, and the MCP project publishes no evaluation. It remains heavily installed, at about 143,000 npm downloads in the week to 18 October 2026, largely on the strength of early 'must-have' lists.
Security exposure is about as low as an MCP server gets: no files, no network, no credentials, and the history vanishes when the connection closes. The one detail to know is that thoughts are written to stderr by default, so whatever the model reasons about, pasted secrets included, can end up in a client's log files.
Where it earns its place
A developer running an 8B local model for coding help
Get the model to plan a multi-step refactor instead of answering in one shot
The numbered-thought schema pushes a small model to list steps and revise them before writing code. It is free and local; each thought is an extra round trip, slow on a laptop GPU but costing nothing.
A compliance lead at a fintech piloting an internal agent
Keep a record of how the agent reasoned before it recommended an action
Each thought is a tool call with its own arguments, so it appears in the client's tool log and the server's stderr log for later review. The trace is what the model chose to write down, not proof of how it actually decided.
An engineer trimming token spend on a production Claude or GPT agent
Decide which MCP servers to drop from a crowded configuration
This is usually the easiest cut: removing it saves about 921 tokens per session plus every thought call, and a reasoning model plans without it.
Before you commit
- Setup
- Add npx -y @modelcontextprotocol/server-sequential-thinking to the client config (README snippets for Claude Desktop and VS Code; Windows needs cmd /c). Node.js 24+ from 2.0.0. It works on restart, and the host decides when to call it.
- Integrations
- One tool, sequentialthinking, with four required inputs (thought, nextThoughtNeeded, thoughtNumber, totalThoughts) and five optional revision and branching fields. It writes only to in-memory history.
- Data and privacy
- No files, network or credentials. Thoughts are held in memory for the connection and logged to stderr unless DISABLE_THOUGHT_LOGGING is set to true.
- Cost at scale
- Free. The cost is tokens: about 921 for the tool definition in every session, per issue #4507, plus a call per thought, so a ten-step plan means ten round trips.
- Limits
- No persistence beyond the connection, no published evidence that it improves results, and the README's tool name does not match the registered one.
What's good
- No files, network or credentials
- Leaves a visible, logged trace of each step
- Can help small non-reasoning models plan
What's not
- About 921 tokens of tool definition per session
- Duplicates built-in reasoning in current models
- README gives the wrong tool name (issue #5204)
Best for
People using smaller or older models who want a visible step-by-step planning trace.
Alternatives
Sequential Thinking (reference server): common questions
- Is Sequential Thinking (reference server) worth it?
- Redundant on today's reasoning models, it now justifies its 921-token overhead only as a visible planning trace or as scaffolding for small models. We rate it 3.3 out of 5.
- How much does Sequential Thinking (reference server) cost?
- Free and open source.
- What are the downsides of Sequential Thinking (reference server)?
- About 921 tokens of tool definition per session; duplicates built-in reasoning in current models; rEADME gives the wrong tool name (issue #5204).
- Who is Sequential Thinking (reference server) best for?
- People using smaller or older models who want a visible step-by-step planning trace.
- How much does Sequential Thinking (reference server) cost for a team or in production?
- Free. The cost is tokens: about 921 for the tool definition in every session, per issue #4507, plus a call per thought, so a ten-step plan means ten round trips.
- Is Sequential Thinking (reference server) safe for company data?
- No files, network or credentials. Thoughts are held in memory for the connection and logged to stderr unless DISABLE_THOUGHT_LOGGING is set to true.
- What are the best Sequential Thinking (reference server) alternatives?
- The closest alternatives we have reviewed are Memory (reference server).
Answers checked Oct 23, 2026.