@modelcontextprotocol/server-sequential-thinking

This server connects to nothing external — what it changes is how the AI thinks: it breaks a hard problem into explicit steps, and crucially lets the model revise an earlier conclusion, branch into an alternative line, and converge again.

It suits problems you can't see the shape of upfront — an architecture decision, or a bug with no obvious cause. Instead of jumping to an answer, the AI lays out its reasoning, so you can see each step and catch where it went wrong.