You do not call it by hand — you connect it and the host calls it repeatedly while it works, which is what makes the reasoning visible as a numbered chain instead of one long answer. The parts that matter are revision and branching: a plan that turns out to rest on a bad assumption gets reopened at the step where the assumption entered, rather than restarted.
A single-tool server that structures step-by-step reasoning: each call carries one thought, its number, the estimated total, and whether another step is needed.
- A problem broken into numbered steps, with the estimated total adjustable as the work goes on
- Revision of an earlier step — isRevision and revisesThought mark which thought is being reconsidered
- Alternative lines of reasoning, tracked with branchFromThought and branchId
- An explicit continue-or-stop signal on every step, so the chain ends when the thinking is done rather than at a fixed length
No account, no key, no network calls. Published on npm as @modelcontextprotocol/server-sequential-thinking and launched through npx; the package ships the mcp-server-sequential-thinking binary.
One command — npx -y @modelcontextprotocol/server-sequential-thinking