By LaunchVault editorial · Official sources checked
LearnHarness is an open-source AI tutor for people who want a study process rather than another answer box. The desktop app organizes lessons around a learning goal, asks you to explain your thinking, and uses practice sessions to revisit the material. A knowledge graph connects ideas between sessions.
The practical decision is whether you want to manage your own model access. The software is free, but you supply a supported account, API connection or model endpoint. That flexibility can suit independent learners; it also means setup, running costs and data handling deserve attention before the first lesson.
What Is LearnHarness, and Who Is It For?
Think of it as a learning workspace that coordinates teaching and practice. You bring a question or goal; the agent explores your starting point and develops a route through the concepts you need. Its lesson map makes that route visible instead of leaving you to invent a sequence of chat prompts.
This is most relevant when a subject needs several study sessions: understanding a mathematical argument, revisiting a programming concept or learning the prerequisites for a technical topic. A quick factual lookup may need less structure. These are suggested uses, not learning outcomes measured by LaunchVault.
Choose this approach if you value model choice and an inspectable study record. A learner who wants an account-only service with no provider configuration may find the initial setup less convenient. The app is not a fixed course, examination service or certificate of competence.
Guided Lessons and Adaptive Practice
The teaching workflow asks for more than a correct option. Predictions, confidence and written reasoning give the agent material to work with. For the learner, the useful habit is to explain why an answer seems right, including the point where the reasoning becomes uncertain.
Practice is a separate activity. The documented flow presents problems and adjusts hints to your responses. This distinction matters: following a worked explanation and retrieving an idea without a prompt are different tasks. Use a lesson to develop an explanation, then practice to see whether you can apply it.
The persistent knowledge graph records propositions and the reasoning that connects them, with links to their source sessions. It provides context for later work. Treat it as a map of recorded material rather than a score proving that you have mastered every node.
LearnHarness Download and Desktop Setup
Start with the official download page linked above, which selects the current release rather than a version-specific installer. Check both the operating system and processor architecture. The inspected packages cover Apple Silicon Macs and x64 Windows and Linux machines; do not assume the Mac package supports Intel hardware.
On Windows, the project currently distributes an unsigned installer, so the operating system may show a publisher warning. Linux users can choose AppImage or a Debian/Ubuntu package; the AppImage needs executable permission. Installation notices are documented by the project, not evidence of an independent security audit.
After installation, configure a connection in Settings → Providers and select the Strong, Cheap and Vision model slots. There are no default model selections. Begin with a narrow goal and a short session before attaching a large reading collection: this makes it easier to assess the workflow and understand usage.
API Keys, Subscriptions or Local Models?
The right connection depends on which account or infrastructure you already have. Supported subscription connections, provider API keys, a separately installed Claude Code CLI and a custom OpenAI-compatible endpoint are distinct routes. Their terms and billing do not become interchangeable simply because they appear in the same app.
For a local model, run the backend separately. LearnHarness connects to it; it does not install a model server or download weights. Supply the complete compatible base URL, including /v1 when your server requires it, then connect and choose model IDs. Confirm the backend's context limits and image capabilities instead of relying on the presence of a model in a selector.
A useful first check is one short text lesson. Add image-based material only after confirming that the selected vision model can process it. Optional speech input and web research introduce additional services, so a local language model alone does not establish an entirely offline study setup.
LearnHarness Pricing: Free Software, Separate Running Costs
There is no LearnHarness subscription or added usage markup in the official offer reviewed here. The app is distributed under AGPL-3.0. This describes the software; model providers, existing subscriptions, optional services and the resources needed to run local models can still have costs.
Budget according to your connection. An API key can incur metered charges, a subscription has account limits, and a CLI follows its own authentication and billing configuration. A custom endpoint's unknown price should not be interpreted as a zero-cost inference service.
The app's cost display estimates supported API usage rather than reconciling every connected account. Check the provider's own usage page when deciding what a session cost. For a sensible trial, use a small topic first, inspect usage and only then expand to longer sessions or background research.
Studying From PDFs, Images and Your Own Notes
Desktop sessions can reference local files through the attachment controls. This gives you a way to discuss your own material, but referencing a path is different from copying every file into a permanent archive. Files read by the agent can contribute content to a request sent to the selected model.
Graph extraction is a separate import workflow for supported text formats and PDFs. The getting-started guide distinguishes this from session attachments: PDF text is converted locally without OCR, then model-based extraction follows. Scanned image-only pages may therefore need another preparation step. Local conversion alone does not make the entire import offline.
For a first study session, choose a short, non-sensitive excerpt and a concrete question about it. For example, ask for help explaining why a particular assumption is needed, rather than requesting mastery of a whole textbook. This is an illustrative approach, not a workflow we have tested in the installed app.
Privacy: Local Records Do Not Mean Local-Only Processing
Chats and learning records are stored on your machine. That gives you control over the stored files, but does not determine where inference occurs. A cloud model receives the context supplied to it; search, CLI integrations and dictation have their own external connections.
The documentation describes API-key files as plaintext protected by file permissions, not an encrypted credential vault. Keep profile folders and diagnostic exports private. Before attaching confidential notes, inspect both the application's file-access boundaries and the data policies of the services you enabled.
A useful privacy check is to list each outbound service separately: teaching model, background models, search and speech input. Decide what each may receive. This is more informative than treating a desktop installer or a local graph as a blanket privacy guarantee.
What to Check Before Choosing This AI Learning Assistant
Evaluate the quality of the questions as well as the fluency of the answers. Does a lesson reveal a gap in your reasoning? Can you solve a related problem later? Those are useful checks for your own trial; an attractive lesson map does not answer them by itself.
Also check practical fit: time spent configuring models, availability of your required platform, and whether you are comfortable maintaining local records. Compare the costs and effort with the study method you already use. An open-source AI tutor offers control, while that control brings configuration decisions.
AI-generated explanations and hints can be mistaken. Keep a textbook, course reference or another reliable source available for checking difficult claims. No score, independent learning-effectiveness result or hands-on rating is assigned in this listing.
Sources and Editorial Scope
LaunchVault independently researched this listing using the official pages and documentation below. We have not installed the app or measured learning outcomes. This directory is not affiliated with the developer; installation and model connections take place through the official project and your chosen services.
Sources: Official website · Desktop downloads · Official repository · Getting started · Custom and local models · Usage and cost accounting · Architecture and file access