Adds two tools to your AI agent: `think` for sequential reasoning steps and `verify` for steel-manning your conclusions before you act on them. The design separates idea generation from evaluation to reduce self-bias, based on MetaCrit research showing 76% accuracy gains. Tool descriptions do most of the work, embedding Socratic questioning and dialectical challenges directly in the prompts. You get step tracking (1/3, 2/3, 3/3), adjustable thought counts, and a verify nudge after the final thinking step. Reach for this when you want structured problem-solving with a built-in adversarial check, especially for decisions where confirming your own reasoning isn't enough. Works over stdio with Claude Desktop, Cursor, Windsurf, and similar clients.
claude mcp add --transport stdio io.github.stabgan-steelmind uvx steelmind