MCP server generates and auto-validates concept maps 100% locally via Ollama
HeadBreakfast6511 · reddit · 2026-09-15
An open-source MCP server runs small models (Qwen2.5:14B / Llama 3.1:8B) via Ollama to turn any text into Mermaid concept maps — 100% local, no cloud, no API keys, no token costs. Any MCP client (e.g. Claude Desktop) can call it directly.
Key design: the model never draws the map directly (that caused hallucinations and broken syntax). Instead it extracts structured data in 3 steps — dense summary → key concepts → relationships — and a separate code engine renders the Mermaid diagram. Before returning, a 10-metric evaluator audits the result (weird cycles? full connectivity? real cross-links vs. one linear chain?) and feeds specific failures back for repair.
Unexpected finding: same corpus, same code, 4 runs — approval rate ranged from 14% to 100%. Pure stochastic variance from the small model. The author added controlled repetition (5 runs per text, fixed temperature and seed) to characterize it, and that analysis became the most interesting part of the project.
Repo: github.com/aldasorolauti-pixel/concept-maps-SLM
More from coding & agent
- Stop reading terminals: build with AI entirely in realtime HTML pages — nptacek · 2026-09-15
- Coding agents must survive a day without saying 'You're right, sorry' before ASI, researcher jokes — andrey_kurenkov · 2026-09-15
- Claude Code's Function Hooks bring Express-style middleware to plugins — nicolascraske · 2026-09-15
- Review agent commits in parallel with zero context via the show-me skill — lucasmeijer · 2026-09-15
- Dolphin creator Eric Hartford unveils With, a systems language for AI infrastructure — QuixiAI · 2026-09-15
- Headroom 1.4 adds motion design, new gauges, and Cursor and Grok support — kieranklaassen · 2026-09-15