An AI tutor that tests you first

LearnHarness calls itself "an agent harness for learning." It's a desktop app where AI agents plan and teach lessons, starting from what you already understand and working toward a goal you pick. Most AI tutors just explain things back at you. This one uses your answers and your reasoning to decide what comes next. It runs on macOS (Apple Silicon), Windows and Linux. You bring your own API key.

How it works

Every lesson goes through four stages. Measure finds where your understanding stops for each idea the lesson depends on. Plan builds a dependency map from that point to your goal, and you can look at the map yourself. If your understanding turns out to be different from what the agent assumed, the lesson changes. Teach goes one step at a time: you make a prediction, get an explanation, then check yourself. Confusions get fixed before anything is built on top of them. Practice happens in separate sessions with varied problems, and the hints adapt to how you're reasoning.

You explain your thinking in "thinking fields," either typed or dictated. The agent can also do background research, make visual explanations and build interactive demos.

The part I like most is the knowledge graph. Lessons pile up as propositions and reasoning chains, each linked to the session it came from. You can import your own notes, and later sessions draw on the graph. The README is explicit that it stores ideas and how they connect, "not ratings of your ability."

Try it

Prebuilt downloads are on the repo page: a signed DMG for macOS, an unsigned installer for Windows (expect a SmartScreen warning), and an AppImage or .deb for Linux. To run it from source:

npm install
npm run dev        # http://localhost:3000
npm test           # deterministic, no LLM, no network

After that, add a provider key in settings and choose your models.

My take

The teaching loop is the real thing here. Measure, then plan, then repair confusions before moving on is how good tutors actually work, and showing you the dependency map makes the plan something you can inspect. I also like that the design is honest. System prompts can be seen and edited. Chats, the graph, settings and keys are stored as plain files on your machine. A live meter shows what the models are costing you, and there's no subscription or markup. The test suite runs with no LLM and no network, which is rare and nice for an agent project.

There are limits. Local storage does not mean offline, and the README says so directly. Model calls, web research and dictation all go to outside services. Dictation and web search each need their own extra key. The Windows build isn't signed yet. How good the lessons are will depend on the model you plug in. The license is AGPL-3.0, so think about that before you fork it into a product.

Who should care: self-learners who want structure instead of a chat window, and devs building agent harnesses who want to see a real multi-stage loop with persistent memory.