Steelmind

io.github.stabganPublished by GitHub account stabganv2.0.1Updated Oct 24, 2026

Structured thinking + steel-manning verification for AI agents. Backed by 43+ papers.

VerifiedSTDIODesktop onlyProductivity & WorkflowAI & ML

Overview

AI-generated overview

A local MCP server that gives an assistant two reasoning tools: recording structured thinking steps and steel-manning a conclusion before acting.

What you can do
Record sequential reasoning steps with adjustable step count, with Socratic self-questioning prompts built into the tool description and a verify nudge on the final step. Challenge a conclusion with a steel-manning prompt that asks for the strongest argument against it. The verify tool returns the concern unchanged; the guidance lives in the tool descriptions rather than in computation.
When to use it
Useful when you want an assistant to break a problem into explicit steps and pause to challenge its own conclusion before committing, for example in planning, analysis or design decisions. Not needed for simple lookups or tasks with no reasoning trade-offs.
Requirements
Runs locally over stdio. Install via npx with the package name, a global npm install, or Docker; Node.js is needed for the npm and npx routes. No accounts, API keys, environment variables or headers are declared.
Before you install
The server only records and echoes text; it does not verify facts or check reasoning itself, so the benefit depends on the model following the tool descriptions. Reasoning steps and concerns are passed to the local process and may appear in conversation logs.

Installation

In SourceWeft

  1. Open Steelmind in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

Steelmind MCP — Structured Thinking & Verification for AI Agents

[npm version] [Docker] [License: MIT]

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.

Steelmind gives your AI agent two tools:

  • think — Record structured reasoning steps with sequential decomposition. Embeds Socratic self-questioning and Polya's problem-solving method.
  • verify — Challenge conclusions with steel-manning before committing. Embeds dialectical evaluation from MetaCrit and SIEV research.

The code is minimal. The descriptions do the heavy lifting — tool descriptions account for ~80% of reasoning improvement per Anthropic τ-bench research.

Why Steelmind?

FeatureThink MCPSequential ThinkingSteelmind
Step tracking✗✓✓
Adjustable step count✗✓✓
Cognitive mode separation✗✗✓
Steel-manning verification✗✗✓
Socratic self-questioning✗✗✓
Research-grounded descriptions✗✗✓
Verify nudge on completion✗✗✓
Tool count112

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.

Quick Start

npx (no install)

json
{  "mcpServers": {    "steelmind": {      "command": "npx",      "args": ["-y", "@stabgan/steelmind-mcp"]    }  }}

Docker

json
{  "mcpServers": {    "steelmind": {      "command": "docker",      "args": ["run", "--rm", "-i", "stabgan/steelmind-mcp"]    }  }}

npm global install

bash
npm install -g @stabgan/steelmind-mcp
json
{  "mcpServers": {    "steelmind": {      "command": "steelmind-mcp"    }  }}

How It Works

The think tool

Records a structured reasoning step with sequential tracking.

Input:

json
{  "thought": "What are the dependencies? Need to check imports before refactoring.",  "thoughtNumber": 1,  "totalThoughts": 3,  "nextThoughtNeeded": true}

Output (mid-sequence):

[Thinking 1/3]
What are the dependencies? Need to check imports before refactoring.

Output (final step — includes verify nudge):

[Thinking 3/3]
My conclusion: use the adapter pattern for backward compatibility.
---Thinking complete. Before acting on this conclusion, use the verify tool to challenge it.

The 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.

The verify tool

Challenges your reasoning with steel-manning before you commit.

Input:

json
{  "concern": "The adapter pattern adds complexity. Is the simpler approach actually better?"}

Output:

The adapter pattern adds complexity. Is the simpler approach actually better?

Pure 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?"

The workflow

think(step 1/3) → think(step 2/3) → think(step 3/3) → [verify nudge] → verify → act                                          ↑                                  adjust totalThoughts if needed

Research Foundation

Steelmind's design is grounded in 43+ research papers. Key findings:

PaperFindingHow Steelmind Uses It
MetaCrit (arxiv 2507.15015)Separating generation from evaluation prevents self-biasTwo separate tools: think (generate) + verify (evaluate)
Anthropic τ-benchOptimized tool descriptions yield 54% improvementDescriptions are the primary scaffold, not code
Think2 (arxiv 2602.18806)Structured metacognition yields 3x self-correctionSequential step tracking + Socratic questioning
SIEV (ICML)Models lose 40+ points under dialectical evaluationSteel-manning prompt in verify description
Scaling TTC (arxiv 2408.03314)Difficulty-adaptive compute improves efficiency 4xAdjustable totalThoughts
EasyTool (NAACL 2025)Concise descriptions outperform verbose ones~100 word descriptions
ToolACE"When NOT to use" improves irrelevance detection 6→84%Negative guidance in both descriptions
Cognitive Foundations (arxiv 2511.16660)External scaffolding improves performance up to 72%Research-grounded cognitive frameworks

Compatible Clients

Works with any MCP-compatible client:

  • Claude Desktop / Claude Code
  • Cursor
  • Windsurf
  • Kiro
  • Cline
  • Any client supporting MCP stdio transport

Compatible Models

Designed for frontier models but works across families:

  • Claude (Opus, Sonnet) — native MCP
  • GPT-5 / GPT-4o / o-series — via MCP adapters
  • Gemini — via MCP adapters
  • DeepSeek — via MCP adapters

Development

bash
npm install          # Install dependenciesnpm run build        # Compile TypeScriptnpm test             # Run 90 testsnpm run lint         # ESLintnpm run format       # Prettiernpm start            # Run the server

License

MIT

Source: README.md at commit 612fdc1

Tools

0
SourceWeft has not started this package in its sandbox yet, so its tools are not known.

Version history

1
  1. v2.0.1LatestSep 16, 2026

Score

50% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 days25 pts

    Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.

  • Auth declared10 pts

    Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.

  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts