DeepTutor: An Agent-Native Learning Workspace Hits v1.7.0
By Indie Kings | October 29, 2026
Updated October 29, 2026: HKUDS/DeepTutor, a learning workspace built around AI agents, shipped version 1.7.0 on October 28 and passed 47,800 GitHub stars. The repository has 224 contributors, an Apache-2.0 license and a paper on arXiv, and it describes itself as lifelong personalized tutoring rather than a chatbot with a textbook attached.
Image: library shelves at Royal Holloway, University of London, via Wikimedia Commons.
One runtime, nine modes
Most study tools pick a lane. A flashcard app makes cards, a PDF chat app answers questions about a document, a quiz generator produces questions and forgets them. DeepTutor's pitch is that all of those are the same activity seen from different angles, so it runs them on one capability runtime with shared session context. Chat, Ask Questions, Quiz, Research, Visualize, Solve, Mastery Path, Immersive Reading and Audio Overview each keep their own loop and pipeline, but they draw on the same knowledge bases, books, notebooks, favorites, question banks and memory.
That sharing is the part worth testing, because the failure mode of a suite like this is nine features that cannot see each other. The README's example is a quiz that pulls from a knowledge base you built for a course, with results landing in the same question bank that Mastery Path grades from, and the tutor remembering which topics you missed. Whether the implementation lives up to the diagram is a separate question, but the architecture is at least aimed at the right problem. Version 1.7.0, released on October 28, adds modular installation, a refreshed learning workspace, block notebooks, narrated video lessons and resumable embedding checkpoints.
Memory you can inspect
The memory design is the most unusual part of the project. DeepTutor splits it into three layers: L1 traces, L2 surface summaries and L3 synthesis. L1 is the raw record of what happened, L2 is a summary attached to a particular surface such as chat or reading, and L3 is a synthesis across surfaces. The README says the Memory Graph links L2 facts back to L1 evidence and links L3 conclusions to the surfaces that produced them, which means a learner can open the memory and see why the system believes something about them.
Personalization systems usually hide that reasoning, and hiding it is convenient for the vendor and bad for the student. If a tutor decides you are weak on linear algebra because of one bad quiz, you should be able to find that inference and correct it. The three layer split also gives the project a defensible privacy story: traces stay local in the runtime home, and the synthesis is a derived artifact that can be inspected and edited rather than an opaque embedding. It is the kind of design that costs engineering time and pays for itself the first time a recommendation is wrong.
The knowledge layer is similarly broad. DeepTutor supports versioned RAG libraries across LlamaIndex, PageIndex, GraphRAG, LightRAG, a remote LightRAG server, a self-hosted WeKnora base, Tencent IMA and MarginNote 4 libraries, Kiwix ZIM archives and a linked Obsidian vault, with pluggable document parsing. That list is long enough to be a liability, and the project acknowledges it by making the engine a per-library choice with recorded provenance instead of a single global setting.
Docker, agents and the install path
Getting started is easier than the feature list suggests. A single Docker command pulls the image, publishes port 3782 on localhost and mounts a named volume for data, and the app configures model providers from Settings. There are also source installs for Python 3.11 and Node, and a documented path where you hand the installation to a terminal capable AI agent, which reads a setup guide, asks for credentials and starts the server without printing keys. That last option is a sign of what the project is: an agent-native workspace whose own documentation assumes an agent will do the setup.
The subagent support follows the same logic. From chat you can consult a live harness such as Claude Code, Codex, Grok CLI, Antigravity, Kimi, opencode, MiMo, Hermes or OpenClaw, or a configured Partner with its own persona and private memory. For a student working through a programming course, that means the tutor can hand a problem to a coding agent and bring the result back into the lesson rather than pretending to compile code itself. The Extensible tools layer adds MCP servers, CLI apps, image, video and voice generation, and community skills from EduHub.
Project card
| Item | Detail |
|---|---|
| Repository | HKUDS/DeepTutor on GitHub |
| What it is | Agent-native learning workspace for tutoring, practice and research |
| Latest release | v1.7.0, October 28, 2026 |
| Created | December 28, 2025 |
| Stack | Python 3.11 and later, Next.js 16 front end |
| License | Apache-2.0 |
| Paper | arXiv 2604.26962 |
| Quick start | Docker image on port 3782, or source install with deeptutor init |
| Knowledge engines | LlamaIndex, PageIndex, GraphRAG, LightRAG, WeKnora, Kiwix, Obsidian |
| Scale | 47,807 stars, 5,937 forks, 224 contributors, 199 open issues |
FAQ
What is DeepTutor?
A self-hosted learning workspace from the HKU data science group that combines chat tutoring, quizzes, research, visualization and mastery practice in one application.
Is it free?
The code is Apache-2.0, but you supply the model provider, so usage costs depend on the API or local model you connect.
Does it work offline?
Parts of it can, with a local model through an OpenAI-compatible server. Many features, including web search and hosted parsing, need network access.
What makes the memory different?
It is split into traces, surface summaries and synthesis, and the Memory Graph links conclusions back to the evidence that produced them.
How do I install it?
Run the published Docker image on port 3782, or install from source with Python 3.11 and Node, then run deeptutor init and deeptutor start.
Why does it have so many stars so quickly?
The repository appeared in December 2025, ships releases weekly, and sits in the busy intersection of AI agents and study tools, which drove rapid attention.
Bottom Line
DeepTutor is ambitious in a way that invites skepticism: nine modes, a dozen knowledge engines and a memory graph is a lot of surface area for a project that is ten months old. What makes it credible is the release cadence and the inspection design. Shipping 1.7.0 this week with modular installation and resumable embeddings suggests a team that is fixing the unglamorous parts, and the memory layers suggest they have thought about why students would trust the thing. If you want a local, model-agnostic study workspace and you are willing to configure a provider, this is currently the most complete open option.
Sources
- GitHub HKUDS/DeepTutor README, release notes, feature list and installation paths, accessed October 29, 2026
- GitHub API metadata for HKUDS/DeepTutor, stars, forks, contributors, issues and latest commit, accessed October 29, 2026
- deeptutor.info documentation and arXiv 2604.26962 links as listed in the repository, accessed October 29, 2026
Related: Every App: A Prompt That Replaces Your SaaS With Open Source | REA: The MCP Server That Reverse Engineers Anything