# Sequential Thinking MCP

> The official MCP reference server for structured reasoning — a sequential_thinking tool that lets agents decompose, revise, and branch their thinking.

One of the seven still-maintained official MCP reference servers, Sequential Thinking exposes a single tool — sequential_thinking — that scaffolds structured problem-solving: numbered thoughts, revisions of earlier ones, branched alternative paths, and dynamic adjustment of how many steps the problem needs. Zero config, no auth, runs locally via npx.

Website: https://github.com/modelcontextprotocol/servers/tree/main/src/sequentialthinking

Sequential Thinking is the reference server that survived the great archiving — one of just seven the MCP project still maintains in `modelcontextprotocol/servers` — and a fixture of "best MCP servers" lists since the protocol launched. It does one thing: gives the model a **structured thinking scaffold** as a tool.

## Highlights

- **One tool, deliberate shape** — `sequential_thinking` records numbered thoughts with `thoughtNumber`/`totalThoughts`, so reasoning is explicit and inspectable.
- **Revision built in** — `isRevision`/`revisesThought` lets the model formally correct an earlier step instead of papering over it.
- **Branching** — `branchFromThought`/`branchId` explores alternative approaches without losing the main line.
- **Dynamic depth** — `needsMoreThoughts` adjusts the plan's length mid-flight when the problem turns out bigger than estimated.
- **Zero-friction** — local stdio via npx or Docker, no account, no auth, MIT/Apache-licensed reference code.

## In an AI-assisted workflow

```bash
claude mcp add sequential-thinking -- npx -y @modelcontextprotocol/server-sequential-thinking
# then:
# > Use sequential thinking to work through this migration plan —
# > revise earlier steps if later constraints invalidate them
```

You don't call the tool; the model does, iteratively. The visible win is on planning-shaped problems — the externalized trail reads like the working notes behind a [task breakdown](/commands/plan/breakdown-task), revisions included.

> [!NOTE]
> Be honest about its 2026 role: frontier models' native extended thinking covers much of what this server pioneered. It remains valuable where you want reasoning **as data** — inspectable, loggable, branchable tool calls — or as scaffolding for weaker models in a mixed fleet. It's also the cleanest first MCP server to study before [building your own](/guides/advanced/building-an-mcp-server): one tool, clear schema, reference-quality code.

## Good to know

Ships from the official `modelcontextprotocol/servers` monorepo (~87k stars) as `@modelcontextprotocol/server-sequential-thinking` on npm. License nuance: the monorepo is transitioning MIT → Apache-2.0 (older contributions stay MIT). The MCP project itself now sits under the Linux Foundation's Agentic AI Foundation — governance context that matters more than it sounds; see [MCP vs A2A](/guides/mcp/mcp-vs-a2a).

---

_Source: https://agentscamp.com/tools/sequential-thinking-mcp — Tool on AgentsCamp._