HKUDS/DeepTutor

DeepTutor is worth a look if you want an LLM-driven study environment that keeps your materials in the loop, rather than another chat window with a tutor prompt. The price is a real setup effort: model-provider configuration, data storage and ongoing model choices.

The problem

A learner typically collects PDFs, textbooks, video lectures and notes, then bounces between a chat assistant, a quiz tool and a note app. Nothing remembers what was already covered, and answers are grounded in whatever text gets pasted in that turn. DeepTutor tries to make the study session itself the organizing unit: knowledge bases, books, notebooks, question banks, personas and Memory are shared across tutoring, solving, quizzing, research and reading modes. The result is one workspace where an explanation can draw on the same material as a later practice session.

What's different

The design centers on one capability runtime shared by Chat, Ask Questions, Quiz, Research, Visualize, Solve, Course Study, Mastery Path, Immersive Reading and Audio Overview, with session context carried between them. Immersive Reading accepts YouTube and Bilibili links and supports transcript search, timestamp-grounded tutoring, saved notes and resumable progress, which is the most concrete bridge between passive video and active questioning. Skills follow an open Agent-Skills folder format, so a skill library is not tied to this project, and EduHub is wired in as the default registry. A CLI exposes the same capabilities as interactive REPL or JSON for other agents.

The trade-off is complexity and coupling. The browser app, CLI, subagents, Partners and multiple modes share one runtime, so understanding what a given mode sees means understanding workspace and account scoping. The project distinguishes a content workspace that agents may read from a private runtime home, and custom workspaces isolate conversations and progress in their own tree; Settings, credentials and Memory remain shared at account level. That scoping is the thing to check before trusting it with personal study material.

Who should try it

Good fit: a self-directed learner with substantial material in several formats who wants quizzes and tutoring grounded in those documents, and who does not mind running a local app and supplying a model provider. It is also a reasonable fit for someone who wants an agent harness to drive study workflows from a terminal or another agent. Poor fit: someone who just wants quick answers in a chat UI, or a course that requires a managed, institutionally administered platform. A shared multi-user deployment is documented, but that is a different operating burden than the single-user desktop trial.

Requirements and limitations

You need Docker or a Python/Node source install, and configuration of model providers in Settings before the tutor does anything useful. The one-command Docker path avoids installing Python or Node, giving a usable trial at port 3782, but a model provider is still required; self-hosting means you pay attention to provider credentials and the data volume that holds settings and local data. The transcripts and video features depend on outside services, with an administrator-configured Invidious instance able to serve YouTube captions and browsing. The README also points users to a reasoning-safety checklist when an answer loses an earlier constraint or cites weak evidence, which is an acknowledgement that citations and consistency need checking rather than a guarantee they will hold.

HotRepo's take

HotRepo's only recorded observation is 2026-10-25, when the repository sat at 42,620 stars with +213 that day on the daily board; there is no comparable daily history, so persistence of interest cannot be judged from these signals. The more useful check before committing is the workspace isolation model: confirm which folders an agent can read and where settings and Memory live. Start with the Docker trial and review the reasoning-safety checklist expectations.