LearnHarness: an A.I. tutor app that tests before it teaches
Added NewCopyable prompt below
Published Last verified Released
Our verdict
The fuller, app-shaped version of the 'make the agent teach me' idea. Where socratic-tutor is one SKILL.md inside your existing agent, LearnHarness is a separate app with a visible lesson plan, separate practice sessions, file attachments and a knowledge graph of what you were taught that later sessions reuse. It is the right pick when you are learning something big over weeks (a new field, a dense paper, a hard library) and want progress that persists; for a quick drill inside your coding agent, the skill is lighter. Independent builders are using it: @yiyibuilds posted a hands-on review (1,054 likes) that liked the live learning graph but said 'at times I felt tested rather than taught', and @hrdk_jain forked it to add pen-tablet input. Limits: one main developer, a month old (first release 17 Oct 2026), unsigned Windows installer, and model calls send your lesson context to whichever provider you configure, so 'local' means local storage, not offline. We have not run it ourselves.
- Who it's for
- Builders and researchers learning a hard new topic over several sessions who want an agent to find their gaps first, and anyone who wants study progress to carry over instead of re-explaining context each chat
- Leverage
- A paced course on any topic from the models you already pay for, with a persistent map of what you have learned that later sessions build on
- Effort
- Low
- Date added
What it does
LearnHarness (amosblomqvist/learnharness, AGPL-3.0) runs four steps. Measure: before teaching, the agent probes each prerequisite with questions and uses your answers, confidence and free-text 'thinking' fields to find where your understanding ends. Plan: it draws a dependency map from there to your goal, which you can see and which adapts as you go. Teach: each step builds on something you already understand; you predict, get an explanation and check yourself, and confusions are repaired before moving on. Practice: separate sessions pose varied problems without naming the concept first. What you were taught is stored as propositions and reasoning chains in a local graph linked to the source sessions, and you can import notes. The desktop app also takes attached PDFs and images, can run background web research (Tavily, Brave, Exa, Firecrawl or Ollama key), draws visual explanations and has dictation. Chats, graph, settings and keys stay on your machine as plain files; system prompts are visible and editable; a live meter estimates API costs.
How builders use it
- Download the app from learnharness.ai or GitHub Releases (signed DMG on Apple Silicon, installer on Windows, AppImage or .deb on Linux).
- Pick a model source: a provider API key, a supported subscription connection, an installed Claude Code CLI, or a custom OpenAI-compatible endpoint such as a local model server.
- Start a session with a concrete goal, attach the paper or docs you are learning from, and use the thinking fields to say why you chose an answer.
- Run separate practice sessions later; they draw problems from the graph so you find out what you can recall without hints.
What to expect
A visible learning plan from your real starting point, fewer gaps papered over, and a knowledge graph you can pick up next week without re-explaining your background.
Copy this: Setup + first session
# Official source: github.com/amosblomqvist/learnharness (AGPL-3.0)
# 1. Download: https://learnharness.ai (or GitHub Releases)
# 2. Settings: choose models (API key, subscription, Claude Code CLI, or OpenAI-compatible endpoint incl. a local server)
# 3. Optional: Settings > Providers > Search for web research (Tavily/Brave/Exa/Firecrawl/Ollama key)
# First session goal (example wording):
I want to understand <topic> well enough to <concrete task>. I already know <background>. Attached: <paper or docs>.
# Run from source instead:
git clone https://github.com/amosblomqvist/learnharness.git && cd learnharness
npm ci && npm run dev # http://localhost:3000Related discoveries
See the newest picks →- ToolResearch & Automation
Open Deep Research (LangChain)
LangChain's MIT deep research agent on LangGraph: plans, researches with parallel sub-agents and writes a report, with the models and search you choose.
Effort: High#Open source
- ToolResearch & Automation
GPT Researcher
An open-source agent that plans research questions, searches the web or your documents and writes a cited report, with the model you choose.
Effort: Medium#Open source
- ToolResearch & Automation
imozuru: research an unfamiliar field down to the original sources
An MIT Claude Code skill that maps a new field, follows citations to the original sources and cites the exact lines.
Effort: MediumPrompt#Plugin / extension#Open source
- ToolResearch & Automation
Gemini Deep Research
Hand off a research question and get back a multi-source report.
Effort: LowPrompt
Get the best new A.I. picks in your inbox
The best new agent workflows and tools we’ve vetted, with a clear verdict on each. Short, occasional, useful.