{
  "markdown": "# Steelmind MCP — Structured Thinking & Verification for AI Agents\n\n[![npm version](https://img.shields.io/npm/v/@stabgan/steelmind-mcp)](https://www.npmjs.com/package/@stabgan/steelmind-mcp)\n[![Docker](https://img.shields.io/docker/v/stabgan/steelmind-mcp?label=docker)](https://hub.docker.com/r/stabgan/steelmind-mcp)\n[![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)\n\n**The research-grounded reasoning MCP server for AI agents.** Combines step-by-step sequential thinking with steel-manning verification — backed by 43+ cognitive science and AI research papers.\n\nSteelmind gives your AI agent two tools:\n\n- **`think`** — Record structured reasoning steps with sequential decomposition. Embeds Socratic self-questioning and Polya's problem-solving method.\n- **`verify`** — Challenge conclusions with steel-manning before committing. Embeds dialectical evaluation from MetaCrit and SIEV research.\n\nThe code is minimal. The descriptions do the heavy lifting — tool descriptions account for ~80% of reasoning improvement per [Anthropic τ-bench research](https://www.anthropic.com/engineering/claude-think-tool).\n\n## Why Steelmind?\n\n| Feature                        | Think MCP | Sequential Thinking | **Steelmind** |\n| ------------------------------ | --------- | ------------------- | ------------- |\n| Step tracking                  | ✗         | ✓                   | ✓             |\n| Adjustable step count          | ✗         | ✓                   | ✓             |\n| Cognitive mode separation      | ✗         | ✗                   | ✓             |\n| Steel-manning verification     | ✗         | ✗                   | ✓             |\n| Socratic self-questioning      | ✗         | ✗                   | ✓             |\n| Research-grounded descriptions | ✗         | ✗                   | ✓             |\n| Verify nudge on completion     | ✗         | ✗                   | ✓             |\n| Tool count                     | 1         | 1                   | 2             |\n\n**Key research insight:** MetaCrit (arxiv 2507.15015) proved that separating reasoning generation from reasoning evaluation prevents self-bias and improves accuracy by up to 76%. Sequential-thinking uses one tool for both. Steelmind separates them.\n\n## Quick Start\n\n### npx (no install)\n\n```json\n{\n  \"mcpServers\": {\n    \"steelmind\": {\n      \"command\": \"npx\",\n      \"args\": [\"-y\", \"@stabgan/steelmind-mcp\"]\n    }\n  }\n}\n```\n\n### Docker\n\n```json\n{\n  \"mcpServers\": {\n    \"steelmind\": {\n      \"command\": \"docker\",\n      \"args\": [\"run\", \"--rm\", \"-i\", \"stabgan/steelmind-mcp\"]\n    }\n  }\n}\n```\n\n### npm global install\n\n```bash\nnpm install -g @stabgan/steelmind-mcp\n```\n\n```json\n{\n  \"mcpServers\": {\n    \"steelmind\": {\n      \"command\": \"steelmind-mcp\"\n    }\n  }\n}\n```\n\n## How It Works\n\n### The `think` tool\n\nRecords a structured reasoning step with sequential tracking.\n\n**Input:**\n\n```json\n{\n  \"thought\": \"What are the dependencies? Need to check imports before refactoring.\",\n  \"thoughtNumber\": 1,\n  \"totalThoughts\": 3,\n  \"nextThoughtNeeded\": true\n}\n```\n\n**Output (mid-sequence):**\n\n```\n[Thinking 1/3]\n\nWhat are the dependencies? Need to check imports before refactoring.\n```\n\n**Output (final step — includes verify nudge):**\n\n```\n[Thinking 3/3]\n\nMy conclusion: use the adapter pattern for backward compatibility.\n\n---\nThinking complete. Before acting on this conclusion, use the verify tool to challenge it.\n```\n\nThe verify nudge appears in the tool result (not just the description), making it far more likely the model will actually call `verify`. Tool results get different attention treatment than descriptions — they're processed as fresh context.\n\n### The `verify` tool\n\nChallenges your reasoning with steel-manning before you commit.\n\n**Input:**\n\n```json\n{\n  \"concern\": \"The adapter pattern adds complexity. Is the simpler approach actually better?\"\n}\n```\n\n**Output:**\n\n```\nThe adapter pattern adds complexity. Is the simpler approach actually better?\n```\n\nPure identity function — returns your concern unchanged. The value is in the description, which prompts: _\"Steel-man the opposition: What is the strongest argument that your conclusion is wrong?\"_\n\n### The workflow\n\n```\nthink(step 1/3) → think(step 2/3) → think(step 3/3) → [verify nudge] → verify → act\n                                          ↑\n                                  adjust totalThoughts if needed\n```\n\n## Research Foundation\n\nSteelmind's design is grounded in 43+ research papers. Key findings:\n\n| Paper                                        | Finding                                                  | How Steelmind Uses It                                    |\n| -------------------------------------------- | -------------------------------------------------------- | -------------------------------------------------------- |\n| **MetaCrit** (arxiv 2507.15015)              | Separating generation from evaluation prevents self-bias | Two separate tools: think (generate) + verify (evaluate) |\n| **Anthropic τ-bench**                        | Optimized tool descriptions yield 54% improvement        | Descriptions are the primary scaffold, not code          |\n| **Think2** (arxiv 2602.18806)                | Structured metacognition yields 3x self-correction       | Sequential step tracking + Socratic questioning          |\n| **SIEV** (ICML)                              | Models lose 40+ points under dialectical evaluation      | Steel-manning prompt in verify description               |\n| **Scaling TTC** (arxiv 2408.03314)           | Difficulty-adaptive compute improves efficiency 4x       | Adjustable totalThoughts                                 |\n| **EasyTool** (NAACL 2025)                    | Concise descriptions outperform verbose ones             | ~100 word descriptions                                   |\n| **ToolACE**                                  | \"When NOT to use\" improves irrelevance detection 6→84%   | Negative guidance in both descriptions                   |\n| **Cognitive Foundations** (arxiv 2511.16660) | External scaffolding improves performance up to 72%      | Research-grounded cognitive frameworks                   |\n\n## Compatible Clients\n\nWorks with any MCP-compatible client:\n\n- Claude Desktop / Claude Code\n- Cursor\n- Windsurf\n- Kiro\n- Cline\n- Any client supporting MCP stdio transport\n\n## Compatible Models\n\nDesigned for frontier models but works across families:\n\n- Claude (Opus, Sonnet) — native MCP\n- GPT-5 / GPT-4o / o-series — via MCP adapters\n- Gemini — via MCP adapters\n- DeepSeek — via MCP adapters\n\n## Development\n\n```bash\nnpm install          # Install dependencies\nnpm run build        # Compile TypeScript\nnpm test             # Run 90 tests\nnpm run lint         # ESLint\nnpm run format       # Prettier\nnpm start            # Run the server\n```\n\n## License\n\nMIT\n",
  "bytes": 6812,
  "sha": "9cbe53fb8b3a1d89d7b92ded2701d570cca76a30579195e109a9e5af036b1b39",
  "repo_slug": "stabgan/steelmind-mcp",
  "fonte": "repo",
  "truncated": false,
  "api": "https://agentalog.com/api/listings/mcp_io_github_stabgan_steelmind_f0e0e484/readme"
}