DeepTutor

• Building a knowledge base from a textbook PDF and asking the tutor to explain a chapter, then quizzing yourself on it
• Setting a goal such as learning linear algebra and working through a Mastery Path that only advances when you pass its checks
• Reading a paper or watching a lecture video with the tutor citing exact pages or timestamps as you ask questions
• Generating a cited Deep Research report on a topic and turning it into a quiz in the same session
• Running a study partner on Telegram, Feishu, or Slack that remembers what you have covered
• Letting Claude Code or Codex call DeepTutor from the terminal to quiz you or research a topic, with machine-readable output

DeepTutor is less a chatbot than a self-hosted study environment: one agent loop drives tutoring, quizzes, research, reading, and a Mastery Path that only moves forward when an assessment says you have learned the point. Two choices stand out. Memory is a set of files you can open and audit, not an opaque profile, and the tutor can be reached from IM apps or driven by other agents. That suits learners and labs who want control over models and data. The cost is operational. You install, configure, and keep updating a fast-moving project with a very large settings surface, and quality depends on the chat and embedding models you pick. Start with the Docker trial and one knowledge base before turning on Partners or multi-user mode.

About DeepTutor

DeepTutor is an open-source learning workspace from the HKUDS group (the Data Intelligence Lab at the University of Hong Kong), first released at the end of 2025 and rebuilt in 2026 around a single agent loop. You install it from PyPI (`pip install deeptutor`), from source, or as one Docker container, run `deeptutor init` to pick a model provider, and open a local web app; a CLI-only build is available for headless use. Every mode runs on the same runtime and session context: Chat (an agent that calls tools such as web, paper, and Zotero search, retrieval over your knowledge bases, GeoGebra analysis, and code execution), Ask Questions, Quiz, Solve for step-by-step reasoning, Deep Research for cited reports, Visualize, Course Study, Audio Overview, Immersive Reading for PDFs, EPUBs, and YouTube or Bilibili videos with page- or timestamp-grounded tutoring, and Mastery Path, which turns a goal into modules of typed knowledge points and only advances when an assessment, or your explicit declaration, says you have mastered a point, with spaced review later. Book compiles knowledge bases and notebooks into an editable living book with quizzes, flash cards, figures, and per-page chat, and Co-Writer is a selection-aware Markdown editor. Knowledge bases can use LlamaIndex, PageIndex, GraphRAG, LightRAG, or a linked Obsidian vault, among other engines, and can follow GitHub repositories or documentation sites. Memory is a three-layer, file-backed system (event traces, per-surface facts, cross-surface synthesis) that you can inspect in a graph rather than a hidden vector store. Beyond the web app, Partners are persistent tutors on the same brain that you can reach over Telegram, Slack, Discord, Feishu, WeCom, WhatsApp, Microsoft Teams, and other IM channels; My Agents lets a chat turn consult a live Claude Code, Codex, or other coding-agent CLI on your machine; and other agents can drive DeepTutor through its CLI's NDJSON output or its FastAPI backend. Optional multi-user mode adds admin, learner, and guardian roles with isolated workspaces. The maintainers state they have no paid online products, so there is no hosted version: you pay only for the model and search providers you connect, or nothing beyond hardware with local models. The tradeoffs are setup and pace. You run and update the service yourself, retrieval needs a separately configured embedding model, the surface is very large, and releases arrive every few days.