What is DeepTutor?
DeepTutor comes from the Data Intelligence Lab at the University of Hong Kong (HKUDS), the group behind LightRAG and nanobot, and is described in an arXiv preprint from April 2026. It is a learning workspace rather than a single chatbot: Chat, Quiz, Research, Solve, Visualize, Mastery Path, Immersive Reading and Video Lesson share one agent loop, session context and a three-layer memory of what the learner has done, stored as readable files.
It is a Python (FastAPI) backend with a Next.js front end, installed from PyPI as deeptutor[full], from source, or as one Docker image. Knowledge bases can use ten retrieval back ends, from LlamaIndex and PageIndex to GraphRAG, LightRAG, Kiwix archives and a linked Obsidian vault. Persistent Partners carry the tutor into 13 chat platforms, from Telegram and Slack to WeCom and Microsoft Teams, and the CLI streams JSON so other agents can drive it.
The repository was created in December 2025, is licensed Apache-2.0, has about 52,700 GitHub stars and is at v1.7.1 (31 October 2026). The README states the team has no paid online product.
Is DeepTutor worth it?
DeepTutor is the most complete self-hosted tutoring system in our reading, and its weak point is the pace that built it. Between 1 September and 31 October 2026 it shipped 15 releases, from v1.6.3 to v1.7.1, after a rewrite of roughly 200,000 lines in April. The result is a workspace that grounds answers in your own documents, generates quizzes and structured books from them, schedules spaced review of missed questions and exports flashcards to Anki.
Two design choices set it apart. Memory is three layers of plain files, raw event traces, curated facts and a cross-subject profile, which an administrator or learner can open, edit and audit in a graph, instead of an opaque vector store. And retrieval is a per-library choice among ten engines, with a deeptutor kb eval command that scores them against your own question set on Recall@k, nDCG and MRR without a model acting as judge. Few rivals let you measure which retrieval works on your material.
The defaults suit one person on a laptop. Authentication is off unless turned on in auth.json, and the docs say that with auth disabled the backend accepts ordinary browser origins. The Docker command binds port 3782 to 127.0.0.1, which is safe locally; publishing it wider without auth exposes every capability, stored API key and code-execution tool. Request screening (rate limit, IP lists) is an optional extra that is also off by default. Everything you ask, plus retrieved passages from your documents, goes to whichever model, embedding, search and speech providers you configure. We found no product telemetry in the repository, and no security advisories are published on GitHub.
The cost is upkeep. v1.6.3 was a breaking front and back end refactor, v1.7.0 split the package into modules and requires uninstalling the old deeptutor-cli package to upgrade, and desktop installers are unsigned previews that expire after seven days. With 137 open issues and 95 open pull requests, a school adopting it needs an owner who tracks releases. Next to OpenMAIC, which turns a topic into a staged lesson with AI teachers, DeepTutor is the better long-term study companion and the heavier system to keep running.
Where it earns its place
A graduate student preparing for qualifying exams from 40 papers and two textbooks
Study from my own reading list with a tutor that remembers my weak spots
Index the PDFs into a knowledge base, ask questions with page-linked citations in Immersive Reading, file missed quiz questions into the Question Bank for spaced review and export flashcards to Anki. The software is free; cost is model and embedding tokens per question and per indexed document.
An IT coordinator at a secondary school piloting AI tutoring with 30 students
Give students a tutor on approved models without handing out API keys
Run the Docker image behind a reverse proxy, enable auth in auth.json, register the admin first, then create Learner accounts with locked capabilities and grant them specific models and knowledge bases; guardians can view reports. Enable the optional request guard, and budget time to test each release before upgrading.
An ML engineer building a RAG pipeline for course material
Decide which retrieval engine actually works on our documents
Write a JSONL file of queries and gold passages, build the same corpus in LlamaIndex, LightRAG and GraphRAG libraries, and compare them with deeptutor kb eval and kb eval-diff. Metrics are computed from rankings, so the scoring itself costs no model calls beyond retrieval.
Before you commit
- Setup
- pip install "deeptutor[full]" on Python 3.11 to 3.14 with Node.js 20.9+, then deeptutor init and deeptutor start to reach 127.0.0.1:3782; or one docker run with a named volume. The bare deeptutor package installs core chat, CLI and SDK only.
- Integrations
- OpenAI, Anthropic, Gemini, xAI, Mistral, NVIDIA NIM, DashScope, local Ollama, LM Studio, llama.cpp and vLLM, plus Codex and GitHub Copilot sign-in; LlamaIndex, PageIndex, GraphRAG, LightRAG, Obsidian, Zotero and Kiwix; MCP services, EduHub and ClawHub skills; Claude Code and Codex as consultable sub-agents; 13 chat platforms for Partners.
- Data and privacy
- Settings, API keys, chats, memory and knowledge bases live in the local data/ directory or Docker volume. Prompts and retrieved document text go to the configured providers. Auth and the request guard are both off by default; skill imports pass a security gate unless --allow-unverified is used. No telemetry described and no published advisories.
- Cost at scale
- No paid edition or hosted service from the maintainers. Spend is model, embedding, search and speech usage per learner, plus a server; hardware requirements are not published, and local parsing or embedding models need more resources.
- Limits
- 137 open issues and 95 open pull requests; breaking refactors in v1.6.3 and v1.7.0; PocketBase integration is single-user only; the request guard does not screen WebSocket traffic or request bodies; Codex OAuth is experimental.
What’s good
- Tutor, quizzes, books, reading and spaced practice in one workspace
- Memory stored as readable, editable files
- Ten retrieval engines, with a built-in retrieval scorer
- Apache-2.0, with a JSON-streaming CLI for other agents
What’s not
- Authentication is off by default
- 15 releases in two months, with breaking upgrades
- Very large surface: many modules, providers and extras to configure
- Desktop app is an unsigned preview
Best for
Self-learners, research groups and schools who want a self-hosted tutor grounded in their own materials and have someone to keep a fast-moving Python and Node app up to date.
DeepTutor alternatives
DeepTutor: common questions
- Is DeepTutor worth it?
- DeepTutor is the broadest open-source tutor you can host yourself, and it changes so fast that whoever runs it should read every release note. We rate it 4.0 out of 5.
- How much does DeepTutor cost?
- Free and open source (Apache-2.0). Self-hosted only; you pay your model, embedding, search and speech providers, or sign in with a ChatGPT plan through the experimental Codex OAuth.
- What are the downsides of DeepTutor?
- Authentication is off by default; 15 releases in two months, with breaking upgrades; very large surface: many modules, providers and extras to configure; desktop app is an unsigned preview.
- Who is DeepTutor best for?
- Self-learners, research groups and schools who want a self-hosted tutor grounded in their own materials and have someone to keep a fast-moving Python and Node app up to date.
- How much does DeepTutor cost for a team or in production?
- No paid edition or hosted service from the maintainers. Spend is model, embedding, search and speech usage per learner, plus a server; hardware requirements are not published, and local parsing or embedding models need more resources.
- Is DeepTutor safe for company data?
- Settings, API keys, chats, memory and knowledge bases live in the local data/ directory or Docker volume. Prompts and retrieved document text go to the configured providers. Auth and the request guard are both off by default; skill imports pass a security gate unless --allow-unverified is used. No telemetry described and no published advisories.
- What are the best DeepTutor alternatives?
- The closest alternatives we have reviewed are OpenMAIC, Khoj and Open Notebook.
Answers checked Oct 31, 2026.
Make DeepTutor? Show this rating on your site
