deep-thinker is an advanced cognitive reasoning MCP server based on a DAG thought graph. Its purpose is to provide structured deep reasoning that goes beyond traditional chain-of-thought sequential thinking through graph-based thought management, pattern validation, and intelligent strategy selection.
Key capabilities include:
- DAG thought graph: Thoughts form a directed acyclic graph, supporting branching, merging, and cross edges, rather than only linear chains.
- 10 reasoning strategies: sequential, dialectical (thesis→antithesis→synthesis), parallel, analogical, abductive, first principles (deconstruct to basic elements), counterfactual (hypotheses and ripple effects), systems thinking (feedback loops and leverage points), MCTS (Monte Carlo optimization), auto (intelligently selected automatically based on content and graph state).
- Node aliases: In any nodeId parameter, "last", "best", "root" can be used instead of obscure node IDs.
- Structured responses: All tool responses return a consistent MCPResponse JSON, including status, summary, confidence, nextSuggested.
- Session persistence: Automatically saves the thought graph to ~/.deep-thinker/sessions/; can be restored across MCP restarts via reset({ resume: "name" }).
- Friendly error messages: Zod validation errors are converted into human-readable hints.
- Confidence scoring: Multi-factor confidence evaluation, including support/contradiction analysis, depth penalty, and knowledge integration bonus.
- Self-critique: Automatically generates critiques, including severity levels and confidence adjustments.
- Metacognitive engine: Detects stuck states, stagnation, and confidence decline; suggests strategy switches and corrective actions.
- Knowledge integration: Attaches external knowledge to thoughts, detects gaps, and validates cross-source consistency.
- Thought pruning: Dead-end detection, redundancy removal, elimination of deep unproductive branches, path optimization.
- help tool: Discovers all 17 tools grouped by category (core/advanced/workflow) and quick-start examples.
- conclude tool: Comprehensive graph summary, including primaryFinding, actionItems, graphHealth, and nextSuggested.
- High-IQ reasoning enhancements: 8 advanced tools, covering visualization, devil's advocate, interdisciplinary synthesis, temporal projection, ethical assessment, emotional intelligence analysis, decision explanation, and social impact analysis.
- Emotional intelligence: Analyzes emotional tone, empathy, persuasive effectiveness, and stakeholder emotions.
- Ethical framework: evaluation through deontological, consequentialist, virtue ethics, and rights-based perspectives.
- Cross-domain synthesis: combining insights from biology, economics, physics, psychology, computer science, and art.
- Temporal reasoning: projecting thinking into future/past scenarios, including optimistic, pessimistic, realistic, and disruptive scenarios.
- Social impact modeling: analyzing stakeholder sentiment, group cohesion, persuasion effectiveness, and ethical consistency.
- Uncertainty quantification: confidence intervals, probability distributions, and sensitivity analysis for robust decision-making.
- Multilingual support: thinking in English, Turkish, German, French, Spanish, Japanese, Chinese, and Russian.
- Metacognitive levels: recursive reasoning across 5 levels of metacognition.
- PromptOptimizer (zero point): entry point that transforms vague prompts into optimized super-prompts and automatically performs strategy routing.
Usage: npx (no installation required). MCP configuration needs to be added to claude_desktop_config.json, or installed globally. The server communicates via stdio; point the MCP client to the deep-thinker command or node path/to/dist/index.js.
Response format: all tool responses follow the MCPResponse structure, and the nextSuggested field always recommends the next logical step, making it easy to chain tool calls.
Session persistence: after each think call, the thought graph is automatically saved, and sessions are stored in ~/.deep-thinker/sessions/.
Strategy details: the auto strategy analyzes content keywords and the current graph state, for example, when it contains "why/neden/how/nasıl" it shifts to abduction, when it contains "if/eğer/what if/varsayalım" it shifts to counterfactual, when it contains "vs/veya/compare/karşılaştır" it shifts to dialectical, when it contains "system/sistem/loop/döngü" it shifts to systems thinking, and when it contains "fundamental/temel/assumption/varsayım" it shifts to first principles; when average confidence is low and there are many nodes, it shifts to parallel, the first thought shifts to sequential, and after 4 or more sequential thoughts it shifts to dialectical. Edge types include derives_from, contradicts, supports, refines, challenges, synthesizes, parallels, abstracts, instantiates.
Applicable scenarios: tasks requiring structured deep reasoning, decision analysis, breaking through thinking deadlocks, interdisciplinary synthesis, ethical and social impact assessment, and uncertainty quantification.
Limitations: the materials do not provide more information about fees, pricing, platform compatibility, performance metrics, or specific usage steps, and none is currently provided.
