An AI tutor that measures before it teaches

What it does

LearnHarness is a desktop app (TypeScript, AGPL-3.0) where AI agents plan and run lessons from your current understanding toward a stated goal. Before teaching, it probes where your knowledge ends, builds a dependency map to the goal, and adapts the plan as your answers reveal gaps. Practice sessions then pose varied problems designed to make the right knowledge surface without being named first.

The interesting bit

The “think out loud” fields are the core mechanism: instead of logging right/wrong answers, the agent reads your reasoning, uncertainties, and reactions in your own words. Everything you learn accumulates as a persistent knowledge graph of propositions and reasoning chains, so future sessions start from your actual mental model rather than a blank slate.

Key highlights

• Four-phase loop: measure, plan, teach, practice — with confusions repaired before anything is built on them.
• Knowledge graph stores ideas and their connections (not ability ratings), with links back to source sessions; notes can be imported.
• Bring your own keys: 30+ providers (OpenAI, Anthropic, Google, DeepSeek, OpenRouter), no subscription or markup, live cost meter.
• Local storage for chats, graph, settings, and keys — plain files you can inspect; system prompts are visible and editable.
• Agent can research questions in the background, create visual explanations, and build interactive demos.

Caveats

• “Local storage” is not offline processing — model, web search, and dictation requests all hit external services.
• The Windows build is currently unsigned, so expect SmartScreen warnings.
• Dictation and web search require separate service keys beyond your model provider key.

Verdict

Worth a look if you want a tutor that adapts to your reasoning rather than dumping explanations at you — and you’re fine paying your own API costs. If you wanted a fully offline learning tool, this isn’t it.