Skip to content
agentscamp
Tool

Sequential Thinking MCP

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

open sourcemcp
Updated Jun 11, 2026
mcpreasoningplanningreference-serverthinking
Visit website Source
View as Markdown

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.

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

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, 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: 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.

Frequently asked questions

What does the Sequential Thinking server actually do?
It gives the model a thinking scaffold as a tool: each call records a numbered thought with fields for revising an earlier thought (isRevision/revisesThought), branching an alternative line (branchFromThought/branchId), and adjusting the estimated total. The structure nudges models into decompose-revise-verify behavior on multi-step problems instead of one-shot answers.
Do I still need it now that models have extended thinking?
Less than in 2024–25, honestly. Native extended thinking covers much of the ground for frontier models. It still earns a slot when you want thinking to be externalized and inspectable as explicit tool calls, when driving models without strong native reasoning, or in multi-agent setups where one agent's reasoning trail is another's input.
How do I add it to Claude Code?
claude mcp add sequential-thinking -- npx -y @modelcontextprotocol/server-sequential-thinking. No API key, no config; set DISABLE_THOUGHT_LOGGING=true if you don't want thoughts echoed to the console.

Related