What it does

Smart-Thinking is a local-first MCP server that performs graph-based, multi-step reasoning without relying on external AI APIs. It builds a connected thought graph, evaluates nodes using deterministic heuristics—TF-IDF similarity scoring, confidence assessment, quality ranking—and maintains a verification ledger to track reasoning decisions and surface uncertainties. Every inference step is logged with structured metadata, making the pipeline auditable and reproducible. Sessions persist across runs, allowing you to resume reasoning with the full graph and verification history intact.

Who it's for

Backend engineers building AI features that need local, auditable reasoning without external API calls. Teams in regulated or air-gapped environments where inference must stay local. Developers using Claude Desktop or Cursor who want to layer deterministic graph-based memory and context retrieval into their reasoning pipeline.

Common use cases

• Build auditable reasoning chains where each decision step is logged, verifiable, and reproducible.
• Extend Claude Desktop or Cursor with deterministic graph-based memory and context retrieval.
• Implement local-only reasoning for compliance-sensitive applications without external API calls.
• Integrate with ChatGPT Connectors in connector mode for knowledge retrieval and verification workflows.

Setup pitfalls

• Requires Node.js and npm; global install via npm install -g smart-thinking-mcp is recommended, but npx fallback is available.
• ChatGPT Connectors and some other clients require --mode=connector flag (or SMART_THINKING_MODE=connector env var) to expose only the search and fetch tools.
• Platform-specific configuration for Windows, macOS, Linux, and Claude Desktop integration is documented in GUIDE_INSTALLATION.md; review before first run.
• Last commit was 106 days ago; verify client compatibility with your MCP version before production deployment.