Labsco
MCP SERVER

Smart-Thinking

by Leghis ↗

Multi-step reasoning as a graph, running entirely on your machine — no embedding service, no external LLM, and a session you can resume and replay.

Reasoning Scaffolds & Agent Workflow Engines
Summary
Deterministic by construction, which is the point — the same input replays to the same graph.

Scoring is heuristic and similarity is TF-IDF, so nothing here depends on a provider being up, a key being valid, or a model version behaving the same way tomorrow. That is also the trade: the judgements are rule-based rather than semantic, so treat the confidence numbers as a consistent yardstick rather than a verdict.

What it is

An MCP server that runs graph-based reasoning deterministically and locally. Thoughts become nodes connected by typed relationships — supports, contradicts, refines, contextual links — and are scored by transparent heuristics rather than an LLM call. Similarity search runs on a local TF-IDF and cosine engine, verification statuses and calculation traces are attached to each node, and sessions persist so a reasoning graph can be resumed across runs.

What you get
  • A reasoning graph with typed relationships between thoughts, rather than a flat chain
  • Local similarity search over memory and graph expansion via TF-IDF and cosine similarity — no third-party embedding service
  • Heuristic quality scoring for confidence, relevance and quality, derived from rules you can inspect instead of a model's opinion
  • A verification workflow with detailed statuses and calculation tracing, propagated across graph connections
  • Sessions that persist and resume, keeping the graph and the verification ledger in sync, with structured metadata per step so runs can be replayed deterministically
  • A connector mode — `--mode=connector` or `SMART_THINKING_MODE=connector` — that narrows the surface to the `search` and `fetch` tools some remote clients require
  • A metadata endpoint for registry scanners at `GET /.well-known/mcp/server-card.json`
Requirements

Nothing — no API key. The default mode is fully local and deterministic: external verification and search connectors are disabled, and no remote LLM or embedding provider is required. Install with `npm install -g smart-thinking-mcp` or run `npx -y smart-thinking-mcp`; the published package is `smart-thinking-mcp` 12.0.1 with a binary of the same name. It speaks stdio for Claude Desktop, Claude Code, Cursor and Cline, and HTTP or SSE for remote clients — `--transport=http --host 0.0.0.0 --port 8000` with the connector mode is the documented remote setup. `SMART_THINKING_ENABLE_EXTERNAL_TOOLS=true` opts into external connectors; with them off, verification suggestions stay local (`executePython`, `executeJavaScript`) and external calls return a local fallback. Windows, macOS and Linux.

Setup effort

One command — npm install -g smart-thinking-mcp