← MCP

Sequential Thinking MCP ServerSequential Thinking MCP Server

Model Context Protocol

推理Inference 本地Local 只读Read-only verified lifecycle: active 核验Checked 2026-10-23

提供单个 sequential_thinking 工具，让 Agent 以可修订、可分支的方式逐步推进复杂问题——它不产生智能，只产生结构。

Provides a single sequential_thinking tool that lets an agent work through a complex problem in a revisable, branchable way — it produces no intelligence, only structure.

适合与不适合When to use it, when not to

让 Agent 用结构化、可修订、可分支的方式推进复杂问题分析：分解复杂问题、可能需要修正的分析、初期范围不明确的任务。

不适合简单任务——它不提升模型能力，只约束输出结构，增加调用轮次而无收益。

Scenarios where a complex problem benefits from being worked through step by step in a revisable, branchable way.

Not a fit for adding capability — it produces structure, not intelligence.

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-10-20

出品属地Origin: 国外出品foreign-made　国外主体（网信办公示名单未见对应备案主体）。foreign entity (no matching filing entity in the CAC registry).
国内使用Direct access: 大陆可直连direct OK　官网仓库在 GitHub，大陆直连波动（同日实测 200/000 混合）；npm / PyPI 安装源实测稳定可达。Hosted on GitHub — direct reachability from mainland fluctuates (200/000 mixed on the same day); npm / PyPI install routes measured stable.
备案状态CAC filing: 未见于备案名单not in filing list：未见于网信办公示名单（2026-10-20 检索 13 批附件，https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm）；该产品未面向大陆公众提供服务，备案不适用。not in the CAC public registry (13 batch attachments searched 2026-10-20, https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm); the product does not serve the mainland public, so filing does not apply.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「大陆可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。GitHub 类站点当日实测有波动，口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

11 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness: 本身不调用模型，是被客户端调用的接口实现。 本 server 不产生智能，只提供结构—— 它约束的是 Agent 调用它的方式（逐步思考并允许修订），真正的推理仍由客户端背后的模型完成。
It calls no model itself; it is an interface implementation called by clients. This server produces no intelligence, only structure — what it constrains is how the Agent calls it (step-by-step thinking with revisions allowed); the actual reasoning is still done by the model behind the client.
运行位置Where it runs: 本地 Node.js 进程，通过 stdio 通信。无文件与网络访问需求。
A local Node.js process communicating over stdio. No file or network access needed.
本地文件Local files: 无任何文件访问能力。不提供读写工具，是本分区权限最窄的一类。
No file-access capability whatsoever. It provides no read/write tools — one of the narrowest-permission entries in this section.
关机后的任务Tasks after shutdown: 不支持。纯计算型工具，无状态，无需后台。
Not supported. A pure-computation tool: stateless, no background needed.
工具与扩展Tools & extensions: 仅 1 个工具 sequential_thinking，

参数包含：

thought
thoughtNumber
totalThoughts
nextThoughtNeeded

以及修订（isRevision / revisesThought）与分支（branchFromThought / branchId）控制。

输出为结构化对象。
A single tool sequential_thinking,

whose parameters include:

thought
thoughtNumber
totalThoughts
nextThoughtNeeded

plus revision (isRevision / revisesThought) and branching (branchFromThought / branchId) controls.

Output is a structured object.
上下文与记忆Context & memory: 无跨会话记忆。每次调用接收完整的思考文本，修订与分支通过参数显式传递，不依赖服务端存储。
No cross-session memory. Each call receives the full thinking text; revisions and branches are passed explicitly via parameters, with no reliance on server-side storage.
权限与限制Permissions & limits: 无任何权限风险：不访问文件、不访问网络、不写任何状态。全部权限就是「能调用一个纯函数」。
No permission risk at all: no file access, no network access, writes no state. Its entire permission set is "being able to call one pure function".
适合什么任务Best for: 让 Agent 用结构化、可修订、可分支的方式推进复杂问题的分析。不适合简单任务（增加调用轮次无收益）。
Lets an Agent advance complex problems in a structured, revisable, branchable way. Not a fit for simple tasks (extra call rounds bring no benefit).
传输方式Transport: 仅 stdio。npx / Docker 本地进程（2026-10-08 README 全文核验）；没有远程传输的官方启动方式。Docker 镜像为 mcp/sequentialthinking。
stdio only. npx / Docker local process (full README verification 2026-10-08); there is no official launch method for a remote transport. The Docker image is mcp/sequentialthinking.
认证机制Authentication: 无独立认证层，但这是本分区唯一无实质权限边界的 server—— 无文件、无网络、无状态，认证与否不影响安全。
No separate authentication layer, but this is the only server in this section with no substantive permission boundary — no files, no network, no state; whether auth exists makes no security difference.
权限范围Permission scope: 无文件与网络访问，不读不写任何外部资源。权限范围为空，是 MCP server 里权限风险最低的一类。可放心在敏感环境使用。
No file or network access; reads and writes no external resources. Its permission scope is empty — the lowest permission-risk class among MCP servers. Safe to use in sensitive environments.

头号误解Top misconception

以为它能让模型变聪明，实际只约束输出结构，推理能力来自模型本身Assuming it makes the model smarter — it only constrains the shape of the output; the reasoning comes from the model itself
以为会记住之前的思考，实际每次调用要自己带上完整上下文Assuming it remembers previous thoughts — each call has to carry the full context itself
以为工具越多越好，本 server 只有一个工具且是有意设计Assuming more tools are better — this server deliberately ships exactly one

未知项清单Known unknowns

是否支持 Streamable HTTP（仓库未声明）
分支是否需要在调用方侧维护完整上下文树
与客户端内置推理能力的重叠程度

Whether Streamable HTTP is supported (not declared in the repo)
Whether branches require the caller to maintain a full context tree
The degree of overlap with clients' built-in reasoning capabilities

证据来源Evidence sources

判断可回到以下一手源复核。本站核验日Every judgement can be rechecked against the primary sources below. Verified on 2026-10-23，内容更新日content updated on 2026-10-08。

类型Type	名称Name	链接Link
docs	MCP · Sequential Thinking ServerMCP · Sequential Thinking Server	https://github.com/modelcontextprotocol/servers/tree/main/src/sequentialthinking
changelog	MCP · Sequentialthinking 目录变更历史MCP · Sequentialthinking directory change history	https://github.com/modelcontextprotocol/servers/commits/main/src/sequentialthinking
changelog	MCP · Servers ReleasesMCP · Servers Releases	https://github.com/modelcontextprotocol/servers/releases
repo	MCP · Servers 仓库 README（生产环境免责声明）MCP · Servers repo README (production-use disclaimer)	https://github.com/modelcontextprotocol/servers

实测记录Measured runs

本档未在真实客户端逐项跑通；工具清单与权限来自官方 README。

实测建议：这类 server 的实测重点在调用轮次与 token 成本，而不是能否完成任务。要对比「用 / 不用」在同一任务上的总 token 与轮次差异。

This site has not completed measured runs. The test protocol is at protocols/mcp/_protocol.md.

Suggested measurement: for this class of server the focus is call rounds and token cost, not whether the task completes. Compare total tokens and rounds "with / without" on the same task.

变更记录说明Change-log notes

本对象无独立 changelog 页，变更记录在目录 commit 历史中：

日期	内容
2026-09-03	声明 zod 运行时依赖
2026-08-28	恢复 `nextThoughtNeeded` 在 inputSchema required 数组中的声明
2026-08-28	从 package.json 读取 server 版本（不再硬编码）
2026-07-29	依赖批量更新

08-28 那条 schema 修复值得注意：nextThoughtNeeded 曾被错误地从 required 数组中移除。如果你的客户端严格校验 schema，这会导致调用被拒。

This object has no standalone changelog page; change records are in the directory's commit history:

Date	Content
2026-09-03	Declare the zod runtime dependency
2026-08-28	Restore `nextThoughtNeeded` in the inputSchema required array
2026-08-28	Read the server version from package.json (no longer hardcoded)
2026-07-29	Batch dependency updates

The 08-28 schema fix is worth noting: nextThoughtNeeded had been wrongly removed from the required array. If your client validates the schema strictly, this causes calls to be rejected.

工具设计Tool design

只有一个工具，参数分四组：

组	参数	作用
进度	`thought` `thoughtNumber` `totalThoughts` `nextThoughtNeeded`	当前这一步是什么、总共几步、是否还需继续
修订	`isRevision` `revisesThought`	承认要改前面的某一步
分支	`branchFromThought` `branchId`	从某一步分出另一条推理路径
追加	`needsMoreThoughts`	中途发现需要更多步骤

「修订」和「分支」是这个 server 的核心价值—— 它让 Agent 在推理过程中回头改，而不是一条路走到底。

There is only one tool; its parameters fall into four groups:

Group	Parameters	Purpose
Progress	`thought` `thoughtNumber` `totalThoughts` `nextThoughtNeeded`	what this step is, how many steps total, whether to continue
Revision	`isRevision` `revisesThought`	acknowledge revising an earlier step
Branching	`branchFromThought` `branchId`	fork another reasoning path from some step
Extension	`needsMoreThoughts`	discover mid-way that more steps are needed

"Revision" and "branching" are this server's core value — they let the Agent go back and change steps mid-reasoning, instead of one path to the end.

关键澄清：它不提升模型能力Key clarification: it does not boost model capability

这是最容易误解的一点。

本 server 不调用模型，不做任何推理。
它约束的是客户端的调用方式，推理仍由模型完成。

它解决的是「结构」问题，不是「智力」问题：

它做的	它不做的
让推理分步可见	让模型更聪明
允许中途修订	提高结论正确率
支持并行分支探索	减少 token 消耗（相反会增加）
让调用方能观察过程	替代思考预算控制

简单任务上用它会增加调用轮次而无收益。 README 自己列的适用场景也限定在：分解复杂问题、可能需要修正的分析、初期范围不明确的任务、需要过滤无关信息的长任务。

This is the most easily misunderstood point.

This server calls no model and does no reasoning.
It constrains how the client calls it; reasoning is still done by the model.

It solves a "structure" problem, not an "intelligence" problem:

What it does	What it does not do
Makes reasoning visible step by step	Make the model smarter
Allows mid-way revision	Improve conclusion accuracy
Supports parallel branch exploration	Reduce token consumption (it increases it instead)
Lets the caller observe the process	Replace thinking-budget control

Using it on simple tasks adds call rounds with no benefit. The applicable scenarios the README itself lists are also limited to: breaking down complex problems, analyses that may need correction, tasks with unclear initial scope, long tasks needing irrelevant information filtered out.

权限评价Permission assessment

MCP server 里权限风险最低的一类，可以放心在敏感环境使用：

检查项	结果
读文件	❌ 无
写文件	❌ 无
访问网络	❌ 无
持久化状态	❌ 无
删除能力	❌ 无

全部权限就是「能调用一个纯函数」。这类 server 值得单独指出—— 它说明 MCP 生态里也有零风险的工具，不全是权限敏感型。

The lowest permission-risk class among MCP servers; safe in sensitive environments:

Check	Result
Read files	❌ none
Write files	❌ none
Network access	❌ none
Persisted state	❌ none
Delete capability	❌ none

Its entire permission set is "being able to call one pure function". This class of server deserves calling out — it shows the MCP ecosystem also has zero-risk tools; not everything is permission-sensitive.