中文版

LearnHarness（GitHub：https://github.com/amosblomqvist/learnharness ，2026-10-16 创建，今晨抓取时已 5,754★）是一个用 AI Agent 当私人导师的开源桌面应用：先测出你对一门课的理解停在哪一步，再画出从起点到目标的依赖图，然后一步一步教——每步先让你预测、再讲解、再检查理解，卡住的地方先修好再往上盖；另有专门的练习会话，用不点名知识点的变式题逼你把知识想起来用。官网 learnharness.ai，AGPL-3.0 开源。

学到的东西会沉淀成一张持久的知识图谱：命题与推理链，带回源会话的链接，之后的学习与练习都能接着它走，不用从头再来。桌面端可附加 PDF、照片等文件给 Agent 读，也能后台联网做研究、生成可视化讲解与互动演示。

下载即用：macOS（Apple Silicon，已签名公证的 DMG）、Windows（x64 安装包）、Linux（AppImage 或 .deb），应用内更新。模型自带：可用各家 API Key、桌面端支持的订阅连接、已装的 Claude Code CLI，或自定义 OpenAI 兼容端点（含本地模型服务）。没有订阅费，也不加价；聊天记录、图谱、设置与 Key 都存在本机，学习记录是可直接查看的纯文本文件。

为什么值得了解

• 对正在求职、需要快速啃下新领域的人，这是“按你的理解程度定制课程”的学习法：先测后教、卡点先修，比直接要一份讲解更接近真人导师
• 知识图谱是亮点：学过的命题与推理链跨会话保留，越学越像你自己的第二大脑，而不是每次重开一局
• 但模型费用自理（按你选的供应商结算），且教学质量取决于你接的模型；Windows 安装包目前未签名，首次安装会弹 SmartScreen 提示

English Version

LearnHarness (GitHub: https://github.com/amosblomqvist/learnharness , created 2026-10-16, 5,754 stars at this morning’s fetch) is an open-source desktop app that turns an AI agent into a personal tutor: it first measures where your understanding ends on each part a lesson depends on, draws a dependency map from that starting point to your goal, then teaches one reasoning step at a time — you predict, get an explanation, and check your understanding, with confusions repaired before anything is built on them — and separate practice sessions pose varied problems that make the knowledge come to mind without being named. Its site is learnharness.ai, AGPL-3.0 licensed.

What you learn accumulates as a persistent knowledge graph of propositions and reasoning chains, linked back to the source sessions, so future teaching and practice pick up where you left off instead of starting from scratch. The desktop app can attach PDFs, photos and other files for the agent to read, research questions in the background, and build visual explanations and interactive demonstrations.

It is ready to download: macOS (Apple Silicon, signed and notarized DMG), Windows (x64 installer) and Linux (AppImage or .deb), with in-app updates. Bring your own models: 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. There is no LearnHarness subscription or markup; your chats, graph, settings and keys live on your machine, and learning records are plain files you can inspect.

Why it’s worth knowing

• For anyone job-hunting and needing to pick up a new field fast, this is a “course customised to your understanding” way to learn: measure first, teach second, repair confusions first — closer to a real tutor than asking for an explanation
• The knowledge graph is the highlight: propositions and reasoning chains carry across sessions, growing into a second brain of your own instead of restarting every time
• Model costs are yours, though (billed by whichever provider you choose), teaching quality depends on the model you connect, and the Windows installer is currently unsigned, so SmartScreen will warn on first install