Self-hosted · Your data stays yours · MIT
It is not just another chat box. It draws up a study plan, teaches it one item at a time, checks whether you actually understood with multiple-choice questions, keeps track of your questions and notes, and lays all of that out in panels you can see. Everything lives in a folder you choose and is uploaded nowhere — except to the model provider you configure yourself.
Why it is different
The problem with studying alongside a general-purpose chatbot is not that its answers are bad. It is that it has no concept of a lesson: it does not know where you are, does not check whether you understood, and cannot hold on to a question that occurred to you halfway through.
| A general chat box | ilearnassist | |
|---|---|---|
| Study plan | None — you decide what to learn and in what order | A numbered, multi-level plan you can tick off and jump around in |
| Pace | Dumps the whole subject on you at once | Usually one item per turn; it waits for you to continue |
| Any checking? | If you say you understood, it believes you | Sets multiple-choice questions, grades them, names your misconceptions — without revealing the answer |
| Your questions | Sink into the message stream, unrecoverable | Notes carry a quote of the original passage and link back to it |
| Where the data is | On someone else’s server | In the folder you picked on first launch |
Features
None of the below needs extra configuration. Pick a data folder on first launch, create an admin account, paste one model provider key — and all of it is available.
A tutor that knows how to take things in order: it finds out what you want to learn, produces a numbered plan, walks through it item by item once you confirm, and sets multiple-choice questions after each point. It teaches in whatever language you write in.
Plan, Quiz, Threads, Notes, Figures, Insights and Resources, in the right-hand panel. They are not decoration — every step of the lesson writes into them, and you can go back at any time.
Highlight a passage, or click a diagram, table, note or resource, and ask about it. The assistant reads that object as it is now, not as it was when it was written.
A persona prompt, a set of allowed tools and a set of default widgets make an assistant. Each conversation copies the definition when it starts, so editing an assistant later never disturbs conversations already underway.
Upload PDF / Word / Excel / PPT / plain text, or keep web pages. Type @ in the composer to pick one as context for the turn — or @ an entire workspace.
PDF and Office files are extracted to text locally, with no network. Scanned documents can optionally go through a cloud parser (MinerU, LlamaParse and similar), but that is a choice, not a requirement.
Bing and DuckDuckGo work with no key; Tavily and SearXNG are also supported. The assistant can keep a page it found useful as a resource, ready to cite later.
Highlighted code, Markdown, Mermaid diagrams, tables, images and PDF all render inside the app — no switching to another program.
Seven widgets
Most AI study tools hide the process inside the message stream. Each widget here maps to one part of a lesson, and fills itself in as the lesson proceeds.
The built-in assistant
It is a public assistant — every account can see and use it — and it has all seven widgets enabled.
Its register is deep, practical and best-practice oriented, not an introductory overview. The language is yours to pick: write in Chinese and it teaches in Chinese, write in English and it teaches in English — switch mid-course and the plan, the list, the questions and the summary all follow. On first use of a technical term it gives the original alongside the translation, to avoid the ambiguity translation introduces.
Privacy and self-hosting
Anyone can say “local-first”. Here are the specifics, each of which you can verify by opening the folder.
You pick the location on first launch (the panel suggests ~/ilearnassist). Inside: db/sqlite/ilearnassist.sqlite (the database), users/<name>/workspaces/ (workspaces and conversation files), users/<name>/sources/ (uploaded material).
No account system, no telemetry, no phone-home. Apart from the one provider whose key you entered, the app initiates nothing — web searches happen only when you ask for them.
The server listens on 127.0.0.1 only. Reaching it from a phone by QR code requires you to switch on a toggle in the control panel; once on, devices on the same network can see your workspaces, conversations and key configuration. That is why it ships off.
The program and the data live in two places. Removing the app leaves your data folder untouched; to erase everything, delete that folder yourself.
When the database schema changes, the app copies it to <data folder>/backups/ before upgrading, keeping the last five — because an upgrade that “succeeds” with the wrong result has no other way back.
DeepSeek, Zhipu GLM, Qwen, Kimi, MiniMax, OpenAI and Google Gemini are built in, and any OpenAI-compatible service works too (Ollama, LM Studio, vLLM, your own gateway). Changing providers does not mean changing tools.
Demos
All six recordings below are real sessions — nothing is cut. Click to play.
Getting started
The installer is the recommended path. If you would rather not use one, the from-source route is below.
Installers are hosted on GitHub Releases. Pick the build for your system and install it as usual.
Go to the download pageThis project has not bought a code-signing certificate, so your system will stop you once. Nothing is wrong — the system simply does not recognise the developer.
Because there is no certificate, there is also no one-click auto-update — on macOS that requires a paid signature. All three platforms work the same way instead (download the new installer and install over the old one; your data lives in the data folder, which is untouched). That is more honest than a button that only works on two platforms.
Requires Node.js ≥ 20 and pnpm ≥ 9.
git clone https://github.com/waychan23/ilearnassist.git
cd ilearnassist
pnpm install
# 配置：数据目录 + 模型 Key
cp .env.example .env
# 编辑 .env，至少填一个 DEEPSEEK_API_KEY（或其他服务商的 Key）
# 全新的数据目录里没有管理员，先创建一个
pnpm --filter @ilearnassist/server cli create-admin --username <你的用户名> --generate
pnpm devIf your firewall asks on first launch, allow the app to communicate on the local machine (127.0.0.1). Nothing is exposed by default.
Model providers
These are seeded on first launch; afterwards the Platform console is authoritative. Domestic and international endpoints are separate entries — the URLs and keys are not interchangeable — so pick whichever suits your network and paste one key.
api.deepseek.com/v1open.bigmodel.cn/api/paas/v4dashscope.aliyuncs.com/compatible-mode/v1api.moonshot.cn/v1api.minimaxi.com/v1api.openai.com/v1generativelanguage.googleapis.com/v1beta/openai/Only configured providers are shown. A provider with no key never appears in the model picker — better than offering a provider that does nothing when you click it.
Any OpenAI-compatible service can be added by hand in the console: Ollama, LM Studio, vLLM, your own gateway…
FAQ
In the data folder you chose on first launch (the panel suggests ~/ilearnassist). The control panel shows you the path and has a button to open it in your file manager. Copying that folder is a complete backup.
Use “Reset admin password” in the desktop control panel. It does not need the server to be running — a forgotten password tends to be discovered alongside other problems, so recovery must not depend on a service that may itself be broken.
No API key has been entered. Configure one under Platform console → Model services and it appears in the model picker immediately.
Yes. The control panel has a LAN sharing toggle that shows a QR code; scan it from a phone on the same Wi-Fi. Think before you switch it on: it exposes your workspaces, conversations and key configuration to other devices on the network. It is off by default.
No. The program and the data live in two places; removing the app leaves your data folder alone. To erase everything, delete that folder yourself.