Deep Thinker

deep-thinker

[](https://github.com/hubinoretros/deep-thinker/actions/workflows/ci.yml) [](https://www.npmjs.com/package/deep-thinker) [](https://www.npmjs.com/package/deep-thinker) [](https://github.com/hubinoretros/deep-thinker/blob/master/LICENSE) [](https://github.com/hubinoretros/deep-thinker/stargazers)

Advanced cognitive thinking MCP server with DAG-based thought graph, 10 reasoning strategies (including auto-selection), 17 tools, node aliases, session persistence, structured responses, and intelligent error handling.

A significant evolution beyond sequential-thinking MCP, providing structured deep reasoning with graph-based thought management, schema validation, and intelligent strategy selection.

Quick Start

npx deep-thinker

{
  "mcpServers": {
    "deep-thinker": {
      "command": "npx",
      "args": ["-y", "deep-thinker"]
    }
  }
}

Examples

| Example | Strategy | Use Case | |---------|----------|----------| | [Architecture Decision](examples/architecture-decision.md) | Dialectic + Parallel | Monolith vs microservices | | [Debugging Incident](examples/debugging-incident.md) | Abductive | Production 500 errors | | [Feature Prioritization](examples/feature-prioritization.md) | Parallel + Dialectic | Q3 roadmap planning | | [Scientific Hypothesis](examples/scientific-hypothesis.md) | Analogical + Abductive | LNP delivery for CRISPR | | [Breaking Dead Ends](examples/breaking-dead-end.md) | Metacognitive switch | Serverless cost analysis |

Features

• DAG-Based Thought Graph — Thoughts form a directed acyclic graph with branching, merging, and cross-edges (not just a linear chain)
• 10 Reasoning Strategies — Sequential, Dialectic (thesis→antithesis→synthesis), Parallel, Analogical, Abductive, First Principles (deconstruct to fundamentals), Counterfactual (what-if with ripple effects), Systems Thinking (feedback loops & leverage points), MCTS (Monte Carlo optimization), Auto (intelligent auto-selection based on content and graph state)
• Node Aliases — Use "last", "best", "root" instead of cryptic node IDs for any nodeId parameter
• Structured Responses — All tool responses return consistent MCPResponse JSON with status, summary, confidence, nextSuggested action
• Session Persistence — Auto-saves thought graph to ~/.deep-thinker/sessions/; resume across MCP restarts with reset({ resume: "name" })
• Friendly Error Messages — Zod validation errors translated to human-readable hints (e.g., "confidence 0 ile 1 arasında...")
• Confidence Scoring — Multi-factor confidence evaluation with support/contradiction analysis, depth penalties, and knowledge integration boosts
• Self-Critique — Automatic critique generation with severity levels and confidence adjustments
• Metacognitive Engine — Detects stuck states, stagnation, declining confidence; suggests strategy switches and corrective actions
• Knowledge Integration — Attach external knowledge to thoughts, detect gaps, validate consistency across sources
• Thought Pruning — Dead-end detection, redundancy removal, deep unproductive branch elimination, path optimization
• help Tool — Discover all 17 tools grouped by category (core/advanced/workflow) with quick-start examples
• conclude Tool — Comprehensive graph summary with primaryFinding, actionItems, graphHealth, and nextSuggested
• High-IQ Reasoning Enhancements — 8 advanced tools: visualization, devil's advocate, cross-disciplinary synthesis, temporal projection, ethical evaluation, emotional intelligence analysis, decision explanation, social impact analysis
• Emotional Intelligence — Analyze emotional tone, empathy, persuasion effectiveness, stakeholder emotions
• Ethical Frameworks — Evaluate through deontological, consequentialist, virtue ethics, rights-based perspectives
• Cross-Domain Synthesis — Combine insights from biology, economics, physics, psychology, computer science, art
• Temporal Reasoning — Project thoughts into future/past scenarios with optimistic, pessimistic, realistic, disruptive scenarios
• Social Impact Modeling — Analyze stakeholder emotions, group cohesion, persuasion effectiveness, ethical alignment
• Uncertainty Quantification — Confidence intervals, probability distributions, sensitivity analysis for robust decisions
• Multi-Language Support — Thoughts in English, Turkish, German, French, Spanish, Japanese, Chinese, Russian
• Meta-Cognitive Layers — Recursive reasoning across 5 levels of meta-cognition
• PromptOptimizer (Node Zero) — Entry point that transforms vague prompts into optimized Super Prompts with automatic strategy routing

Installation

Global

npm install -g deep-thinker

npx (no install)

npx deep-thinker

MCP Configuration

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "deep-thinker": {
      "command": "npx",
      "args": ["-y", "deep-thinker"]
    }
  }
}

Or if installed globally:

{
  "mcpServers": {
    "deep-thinker": {
      "command": "deep-thinker"
    }
  }
}

Other MCP Clients

The server communicates over stdio. Point your MCP client to the deep-thinker command or node path/to/dist/index.js.

Response Format

All tool responses follow the MCPResponse structure:

{
  "status": "ok | error | warning",
  "nodeId": "thought_3",
  "summary": "sequential stratejisiyle \"Should we use microservices?...\" eklendi",
  "confidence": 0.75,
  "data": { "...": "tool-specific data" },
  "nextSuggested": {
    "tool": "evaluate",
    "params": { "critique": true },
    "reason": "Düşük confidence — değerlendirme önerilir"
  },
  "warnings": ["Stuck detected: ..."]
}

The nextSuggested field always recommends the next logical step, making it easy to chain tool calls without guessing.

Node Aliases

Instead of looking up cryptic node IDs, use aliases for any nodeId, parentId, or targetId parameter:

| Alias | Resolves To | |-------|-------------| | "last" | Most recently added node (insertion order) | | "best" | Node with highest confidence score | | "root" | First node with no incoming edges |

evaluate({ nodeId: "last" })                              → evaluates the latest thought
simulate_devils_advocate({ nodeId: "best", depth: 2 })    → challenges the strongest thought
graph({ action: "path", nodeId: "root", targetId: "best" }) → traces from root to best conclusion

Session Persistence

Thought graphs are automatically saved after every think call. Sessions are stored in ~/.deep-thinker/sessions/.

// Save current session explicitly
reset({ save: true, saveName: "my-analysis" })

// List saved sessions
reset({ listSessions: true })

// Resume a saved session after MCP restart
reset({ resume: "my-analysis" })

Tools

Core Tools

think

Add a thought to the cognitive graph using a reasoning strategy.

Parameters:

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | content | string | Yes | The thought content | | type | string | No | Thought type: hypothesis, analysis, evidence, conclusion, question, assumption, insight, critique, synthesis, observation | | strategy | string | No | Strategy: sequential, dialectic, parallel, analogical, abductive, first_principles, counterfactual, systems_thinking, mcts, auto | | confidence | number | No | Initial confidence 0-1 (default: 0.5) | | parentId | string | No | Parent node ID or alias (default: last leaf) | | branch | string | No | Branch name for parallel exploration | | tags | string[] | No | Tags for categorization | | edgeTo | object | No | Explicit edge: { targetId, type } | | dialectic | object | No | Dialectic mode: { thesis, antithesis?, synthesis? } | | parallel | array | No | Parallel mode: [{ content, type, confidence }] | | analogical | object | No | Analogical mode: { sourceDomain, mapping, projectedConclusion } | | abductive | object | No | Abductive mode: { observation, explanations[], bestExplanation? } | | firstPrinciples | object | No | First Principles mode: { problem, assumptions?, depth?, domain? } | | counterfactual | object | No | Counterfactual mode: { currentState?, variablesToChange, rippleDepth? } | | systemsThinking | object | No | Systems Thinking mode: { systemDescription?, components, focusArea? } | | mcts | object | No | MCTS mode: { problem?, possibleActions, numSimulations? } | | knowledge | object | No | Attach knowledge: { source, content, relevance } |

Strategy details:

| Strategy | Description | Best For | |----------|-------------|----------| | Sequential | Linear chain: each thought derives from the previous | Step-by-step reasoning | | Dialectic | Thesis → Antithesis → Synthesis pattern to resolve contradictions | Resolving conflicts | | Parallel | Explore multiple independent branches simultaneously | Brainstorming options | | Analogical | Map patterns from a known domain to the current problem | Cross-domain insights | | Abductive | Generate hypotheses and infer the best explanation | Root cause analysis | | First Principles | Deconstruct to fundamental truths, challenge assumptions | Breaking conventions | | Counterfactual | "What-if" scenarios with multi-stage ripple effects | Risk/impact analysis | | Systems Thinking | Feedback loops, leverage points, emergent properties | Complex systems | | MCTS | Monte Carlo Tree Search for optimal decision selection | Optimization problems | | Auto | Automatically selects strategy based on content signals and graph context | Hands-off reasoning |

How auto strategy works:

The auto strategy analyzes your content for keywords and the current graph state:

• Content with "why"/"neden"/"how"/"nasıl" → abductive
• Content with "if"/"eğer"/"what if"/"varsayalım" → counterfactual
• Content with "vs"/"veya"/"compare"/"karşılaştır" → dialectic
• Content with "system"/"sistem"/"loop"/"döngü" → systems_thinking
• Content with "fundamental"/"temel"/"assumption"/"varsayım" → first_principles
• Low avg confidence + many nodes → parallel (break through impasse)
• First thought → sequential
• After 4+ sequential thoughts → dialectic (introduce opposing view)
• Default → sequential

Edge types: derives_from, contradicts, supports, refines, challenges, synthesizes, parallels, abstracts, instantiates

evaluate

Evaluate the thinking process with confidence scoring, critique, and graph health analysis.

Parameters:

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | nodeId | string | No | Specific node to evaluate (accepts aliases: last, best, root) | | critique | boolean | No | Generate self-critique (default: true) | | findGaps | boolean | No | Find knowledge gaps (default: false) | | validateKnowledge | boolean | No | Validate knowledge consistency (default: false) |

metacog

Metacognitive operations — monitor and control the thinking process.

Parameters:

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | action | string | Yes | report = full state, switch = change strategy, auto_update = let system analyze | | strategy | string | No | New strategy (for switch action) | | reason | string | No | Reason for switching (for switch action) |

The metacognitive engine automatically:

• Detects stagnation (confidence not improving)
• Detects declining confidence trends
• Detects excessive contradictions
• Suggests strategy switches, pruning, backtracking, or concluding

graph

Query and visualize the thought graph.

Parameters:

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | action | string | Yes | visualize, stats, path, node, branches, best_path, leaves | | nodeId | string | No | Node ID or alias (for path, node actions) | | targetId | string | No | Target ID or alias (for path action) |

prune

Prune and optimize the thought graph.

Parameters:

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | action | string | Yes | analyze (report only), prune (execute), optimize_path, prune_node | | nodeId | string | No | Node to prune — accepts aliases (for prune_node) | | reason | string | No | Reason (for prune_node) |

reset

Reset the thought graph and start a fresh session, save, or resume a saved session.

Parameters:

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | problem | string | No | New problem statement | | save | boolean | No | Save current session before resetting (default: false) | | saveName | string | No | Name for saved session (recommended if save: true) | | resume | string | No | Resume a previously saved session by name | | listSessions | boolean | No | List all saved sessions |

conclude

Analyze the entire thought graph and produce a comprehensive summary-conclusion with action items and graph health report.

Parameters:

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | detailLevel | string | No | brief, detailed, technical (default: detailed) | | includeCounterfactuals | boolean | No | Include counterfactual analysis (default: false) | | format | string | No | prose, structured, executive (default: structured) |

Response includes:

• primaryFinding — Top conclusion with confidence
• supportingEvidence — Additional high-confidence nodes
• strategiesUsed — Which strategies contributed
• keyInsights — Insight-type nodes from the best path
• actionItems — Prioritized actions derived from conclusions
• graphHealth — Node count, dead ends, avg confidence, recommendation
• nextSuggested — Logical next step (prune if unhealthy, save if done)

help

Discover deep-thinker tools and learn usage workflows.

Parameters:

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | category | string | No | all, core, advanced, workflow (default: all) |

Categories:

• core — 7 daily-use tools (think, evaluate, metacog, graph, prune, reset, conclude)
• advanced — 8 deep-analysis tools (visualization, devil's advocate, cross-disciplinary, temporal, ethical, emotional, explanation, social impact, prompt optimizer)
• workflow — 3 recommended workflows:
• Quick Decision — reset → think parallel → evaluate → conclude
• Deep Analysis — reset → first_principles → counterfactual → devil's advocate → evaluate → metacog → prune → conclude
• Breaking Dead Ends — metacog report → switch strategy → cross-disciplinary → abductive

Enhanced Tools (High-IQ Reasoning)

visualize_thought_graph

Generate visual representation of the thought graph as SVG or ASCII.

Parameters:

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | format | string | No | svg, ascii, or tree (default: ascii) | | highlightPath | string | No | Path between two node IDs (format: fromId-toId) | | showConfidence | boolean | No | Show confidence scores (default: true) |

simulate_devils_advocate

Generate counterarguments and opposing viewpoints for a given thought.

Parameters:

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | nodeId | string | Yes | Target node ID or alias (last, best, root) | | depth | number | No | Levels of counterarguments (1-5, default: 2) | | intensity | string | No | mild, moderate, or `aggr

…

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: nachosystems
• Source: nachosystems/deep-thinker
• License: MIT
• Homepage: https://www.npmjs.com/package/deep-thinker

Install and usage instructions live in the source repository linked above.