LearnHarness

Overview

LearnHarness is an open-source AI learning application that uses agents to plan and teach lessons from a learner’s current understanding to a goal. It measures understanding, builds a dependency map, teaches step by step, and supports practice with a persistent knowledge graph.

Key Features

• AI agents measure current understanding before teaching and use answers, confidence, and reasoning to decide where to start.
• The agent plans a dependency map from the starting point to the learner’s goal, and the lesson adapts when understanding turns out differently.
• Teaching builds one reasoning step at a time: learners predict, receive an explanation, and check understanding, with confusions repaired before moving on.
• Practice sessions pose varied problems and adapt problems and hints to the learner’s reasoning.
• A persistent knowledge graph stores propositions and reasoning chains, links back to source sessions, and can carry learning across future teaching and practice sessions.
• The desktop app supports attaching PDFs, photos, or other files by path, and the agent can research questions, create visual explanations, mark up images, and build interactive demonstrations.

Use Cases

• Learning difficult topics from a current level of understanding toward a specific goal.
• Finding gaps where a learner may think they understand a topic but cannot solve a new problem.
• Carrying learning and review across sessions through a persistent knowledge graph.
• Studying from PDFs, photos, notes, and interactive demonstrations in a desktop learning session.

Getting Started

• Download LearnHarness from learnharness.ai or GitHub Releases.
• On macOS Apple Silicon, use the signed and notarized DMG and drag the app to Applications.
• On Windows x64, run the installer.
• On Linux x64, make the AppImage executable and open it, or install the Debian/Ubuntu .deb.
• Choose models, then start a session with what you want to understand.

Deployment & Requirements

• A macOS (Apple Silicon), Windows (x64), or Linux (x64) desktop environment.
• A configured model provider: provider API keys, supported subscription connections in the desktop app, an installed Claude Code CLI, or a custom OpenAI-compatible endpoint, including a local model server.
• Optional dictation and web search use separate service keys.
• Web research can use a Tavily, Brave, Exa, Firecrawl, or Ollama key in Settings → Providers → Search; Ollama search uses its hosted service, not a local model server.

Before You Adopt

• License: AGPL-3.0. Review its terms before using, modifying, or distributing the project.
• The Windows installer is currently unsigned, so Windows may display an unknown publisher or SmartScreen warning.
• The live cost meter only estimates known API costs and does not cover subscription, CLI, or custom-endpoint charges.
• Local storage does not mean offline processing, because model requests, web research, and dictation can contact external services.