DeepTutor Review: HKU's Free Open-Source AI Tutor, 52K Stars Strong

Most AI tutors are just chatbots with a school logo. DeepTutor is something different: a full learning workspace that runs on your own computer, keeps track of what you have studied, quizzes you on it later, and builds study materials from your own documents. It is open source under the Apache 2.0 license, it costs nothing to use, and more than 52,000 people have starred it on GitHub since it launched in December 2025.

The project comes from HKUDS, the Data Intelligence Lab at the University of Hong Kong, and it just shipped version 1.7.1 on November 1, 2026. This review covers what DeepTutor actually does, how its memory system works, the four ways to install it, what "free" really means here, and where it still falls short.

What Is DeepTutor?

DeepTutor is an agent-native, open-source personalized tutoring application developed by HKUDS at the University of Hong Kong. The team describes it as "agent-native" because a single agent loop powers everything: the mode you pick — chat, quiz, research, a guided study path — changes only the goal of that session, not the underlying engine.

The numbers tell part of the story. The GitHub repository at HKUDS/DeepTutor holds 52,818 stars and 6,464 forks as of this writing, released under Apache 2.0. The project launched on December 28, 2025, and the accompanying research paper, "DeepTutor: Towards Agentic Personalized Tutoring," is published on arXiv at arXiv:2604.26962. According to the paper as reported by independent coverage, the team's hybrid personalization engine — static knowledge grounding plus dynamic personal memory — improved personalization scores by an average of 10.8% on TutorBench, a 270-task evaluation across five disciplines. That figure comes from the authors' own evaluation, not an independent benchmark.

One Workspace, Eight Learning Modes

The official site at deeptutor.info frames the app as eight connected places sharing one learning flow: Workspaces, Chat, Calendar, Notebook, Learning, Resources, Partners, and Personalization. Workspaces keep each subject's conversations, materials, and progress together, so your Python notes never mix with your history exam prep.

The learning modes include Chat, Quiz, Research, Visualize, Solve, Mastery Path, Reading, and Course, plus writing tools like Co-Writer and Book generation, a Task Board, and a calendar. A few highlights from the current release:

• Mastery Path builds a learning path for a topic and gates progress: you stay on a step until you demonstrate mastery, or you can mark a point mastered yourself.
• Reading opens PDFs and EPUBs beside the tutor so you can read and ask questions side by side, with annotations and text-to-speech for read-aloud.
• Video lessons (introduced in v1.7.0) generate narrated video lessons; v1.7.1 adds checked video lessons, meaning the lessons are verified before delivery.
• Practice and review generate quizzes, store items in a Question Bank, and schedule reviews, with Anki and calendar exports added in the latest release.
• YouTube learning studies videos with synchronized captions, timestamp-linked explanations, saved markers, and resumable playback.

The Memory System Is the Real Differentiator

A normal chatbot forgets you between sessions. DeepTutor's selling point is a three-layer memory that is visible and editable:

• L1 — traces: raw records of what you did in a session.
• L2 — facts: extracted facts about your goals, level, and preferences.
• L3 — synthesis: higher-level summaries of how you learn.

Every layer can be inspected in a Memory Graph, with evidence links you can audit. The idea is that personalization accumulates over months of use instead of resetting every conversation. An independent hands-on review by AI Parabellum, testing version 1.7.0 on October 29, 2026, confirmed the app kept conversation history and suggested a follow-up topic unprompted.

Study From Your Own Documents, Not Just Its Training Data

DeepTutor lets you build versioned knowledge bases from uploaded files, GitHub repositories, and documentation websites. The README lists multiple pluggable RAG engines — LlamaIndex, PageIndex, GraphRAG, and first-party LightRAG — with configurable document parsers including MinerU and Apache Tika, and sources like Obsidian vaults and Kiwix archives.

When you ask a question grounded in your material, answers come with source-linked citations, so you can trace a claim back to the page it came from. Reading mode keeps the document open beside the thread and lets you open a "Little Tutor" thread on a selected passage for a focused explanation. This matters for studying: the tutor teaches from your textbook, not from a vague memory of one.

Partners: Your Coding Agents Can Join the Study Session

One unusual feature is Partners — persistent companions with a persona, private conversations, their own memory, and IM channels. DeepTutor can also consult live agents such as Claude Code, Codex, and the Grok CLI during a tutoring session, or import past conversations from them. If you are stuck on a coding problem, you can bring a coding agent into the study session rather than switching apps.

The project is also extensible by design: personas, community skills from EduHub, MCP services, CLI apps, and generation models plug into the same runtime. Every skill imported from EduHub passes through a registry safety check — defensive archive extraction and content screening — before it can run, which is a sensible precaution for an app that executes tools on your machine.

How to Install DeepTutor: Four Paths

The official documentation at docs.deeptutor.info offers four installation paths. All of them land at the same web app, served on your machine at http://127.0.0.1:3782.

1. Docker — the fastest full-app trial

No Python or Node.js installation needed:

docker run --rm --name deeptutor \
  -p 127.0.0.1:3782:3782 \
  -v deeptutor-data:/app/data \
  ghcr.io/hkuds/deeptutor:latest

Then open http://127.0.0.1:3782. A named volume keeps your settings and data.

2. PyPI — full web app and CLI

Requires Python 3.11–3.14 and Node.js 20.9+ (Node 22 LTS is recommended for the current dev source; narrated video lessons need Node.js 22 or newer):

mkdir -p my-deeptutor && cd my-deeptutor
python -m pip install -U "deeptutor[full]"
deeptutor init     # configure ports, LLM provider, embedding/search
deeptutor start    # starts backend + frontend

Open the URL printed in the terminal. Press Ctrl+C to stop both services.

3. Desktop preview — macOS and Windows

A bundled app, "Open DeepTutor," ships Python, Node.js, and the workspace together — no terminal needed. One caveat from the README: desktop previews are built from the dev branch via GitHub Actions artifacts (GitHub sign-in required, artifacts expire after seven days), and the v1.7.0 release does not include desktop installers. Grab the macos-arm64, macos-x64, or windows-x64 target; Windows builds are unsigned and may trigger a SmartScreen warning.

4. Source — for contributors

Clone the dev branch, create a virtualenv, install requirements, then deeptutor init and deeptutor start. Full commands are in the repo's Get Started section.

After starting the app, you configure a provider and model under Settings → Providers and Language models. That is where the cost question begins.

Pricing: What "Free" Actually Means

DeepTutor's README is blunt: the project iterates "in a fully open-source form" and "we DO NOT have paid online products of any form." There is no subscription, no premium tier, no paywalled feature. In USD, GBP, and CAD alike, the price is $0 / £0 / C$0.

The fine print is the model. DeepTutor is the app; the AI model you connect does the thinking, and that model has its own bill:

• Local models — $0: point DeepTutor at a local model through Ollama or LM Studio and nothing leaves your machine. The whole setup costs nothing.
• Cloud APIs — provider rates: bring your own key and pay whatever the provider charges (for example, Anthropic's or OpenAI's published per-token rates). Embeddings and search calls add to the bill too.

Two honest caveats. First, quality follows the model: the AI Parabellum hands-on test used a small local model (Llama 3.2 3B via Ollama) and it made math mistakes in two of three test problems — the app worked fine, the small model did not. Second, some features send data out: the YouTube-lesson narration uses Edge TTS, which sends lesson text to Microsoft, and web search goes through external providers. For fully private files, the project's own docs recommend local models for both chat and embeddings, plus web search and Edge narration turned off.

What It Is Up Against

DeepTutor occupies a middle ground between plain chatbots and fully local AI apps. A quick comparison:

Nobody else in the free tier combines self-hosted tutoring, visible memory, and document-grounded practice — which is plausibly why the repo is accelerating. Independent tracker whatstrending.ai measured DeepTutor gaining 2,282 stars in a single day on October 31, and it ranked among the week's fastest-growing AI repos. Star counts measure attention, not quality — but a week-over-week acceleration curve is a fair signal that developers find it useful.

Pro Tips and Common Mistakes

Start with Docker if you just want to look. The one-container trial takes minutes and wipes cleanly; reserve the PyPI or source install for when you have committed to a provider and a knowledge base.

Pick your model before your study plan. A 3B local model is fine for organizing notes but will fumble math and science explanations, as independent testing showed. For serious tutoring, use a capable cloud model or a larger local one — and budget the API cost honestly.

Keep subjects in separate workspaces. The workspace model exists so contexts do not bleed together; one workspace per course or project keeps the memory layers clean and the knowledge bases relevant.

Turn off web search and Edge narration for private material. The project's own guidance: point both chat and embedding providers at local models for sensitive documents.

Do not skip the embedding model. The embedding model is what reads your document chunks — if it is weak or misconfigured, retrieval quality suffers no matter how good your chat model is.

Frequently Asked Questions

Is DeepTutor really free?

Yes. It is Apache 2.0 open source, and the maintainers state they have no paid products of any kind. You pay only for the AI model you connect — nothing at all if you use a free local model.

Does DeepTutor work offline?

Yes, if you pair it with local models. Point both the chat model and the embedding model at Ollama or LM Studio, disable web search and Edge narration, and nothing leaves your machine.

Is DeepTutor only for students?

No. It works for any self-learner — the modes cover research, visualization, coding consultation, and book authoring, not just exam prep. Educators can also build lessons and quizzes with it.

How is DeepTutor different from just asking ChatGPT to tutor me?

Three things: persistent, inspectable memory across sessions instead of a reset every chat; document-grounded answers with source citations; and purpose-built practice modes like quizzes, mastery paths, and reading sessions. A chatbot can imitate tutoring; DeepTutor is built around it.

Which models does DeepTutor support?

Anything you can configure as a provider: cloud APIs via your own keys, or local models through Ollama or LM Studio. One vendor key can be linked to every service that provider offers, per the project's setup guides.

Who makes DeepTutor?

HKUDS, the Data Intelligence Lab at the University of Hong Kong, led by Bingxi Zhao, developed in the open with community contributors. The team also runs a collaboration page for institutions that want to work with them.

DeepTutor is the rare open-source project that is both ambitious and honest about its limits: the app is free and the memory system is fully inspectable, but it is only as smart as the model you connect, and the project is clear that no paid support tier exists if you get stuck. If you study from your own materials and want the tutor to remember what you learned last month, it is worth the install. Just bring a capable model.

Sources

• DeepTutor GitHub repository (HKUDS) — github.com/HKUDS/DeepTutor — README read in full; 52,818 stars, 6,464 forks, Apache 2.0, verified November 1, 2026
• DeepTutor official site — deeptutor.info — features and install paths, verified November 1, 2026
• DeepTutor releases — github.com/HKUDS/DeepTutor/releases — v1.7.1 (November 1, 2026) and v1.7.0 (October 28, 2026) notes
• DeepTutor documentation — docs.deeptutor.info — install guides
• DeepTutor paper — arXiv:2604.26962, "DeepTutor: Towards Agentic Personalized Tutoring"
• AI Parabellum: "DeepTutor: How to Install and Use the Free AI Tutor" — hands-on test, October 29, 2026
• XiaoHu.AI: "HKUDS Open-Sources AI Tutor DeepTutor" — paper and feature analysis
• ScriptByAI: "DeepTutor: Open-Source, Multi-agent AI Learning Assistant" — feature overview
• Hands-On PM (pmer.cn): "DeepTutor: lifelong personalized AI tutoring runtime" — review dated October 31, 2026

Related reading: our reviews of Opera WebAI's free local AI, Perplexity Computer on Mac, Claude Code, and Recursive Flint's free 40B coding model.