HKUDS team's agentic learning workbench: It integrates Q&A, problem solving, question generation, research, visualization, close reading, and mastery practice into a single runtime, with the knowledge base and memory shared across all functions.
Stars 45929 (+1598 / 24h)
Reasons
- The knowledge base supports multiple engines including LlamaIndex, PageIndex, GraphRAG, LightRAG, Obsidian, and Kiwix, with version management
- In Reading mode, you can add YouTube or Bilibili links, search by subtitles, ask questions based on timestamps, and take notes
- In chat, you can invoke agents such as Claude Code, Codex, Grok CLI, Kimi, and DeepSeek as sub-agents; the three-layer memory (L1 traces, L2 summaries, L3 synthesis) is viewable and editable
- The latest stable version v1.6.15 (released on 10 19) added learning path review, grouped assessments, Gemini native voice, and more; there is an arXiv paper 2604.26962
Usage
- Quickest experience:
docker run --rm -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 and configure models in Settings → Providers - Local installation:
pip install -U deeptutor, then rundeeptutor initto select the port and model service, anddeeptutor startto launch (requires Python 3.11–3.14 and Node.js 20+) - Alternatively, you can give the prompt provided in the README to Codex / Claude Code, letting the agent set it up according to docs-for-user/AGENT_SETUP.md
Risks
- Requires your own large model API Key; knowledge base retrieval additionally requires an embedding model, and web search requires a search service
- The feature set is broad and updates are frequent (almost weekly), so watch out for breaking changes when upgrading; license is Apache-2.0