DeepTutor puts learning on the same agent loop as everything else
34.7k stars, +829 today, out of HKUDS — the Hong Kong University lab that keeps shipping things people actually run. DeepTutor is an agent-native learning workspace, and the design choice that makes it interesting is that chat, quiz, research, visualization, problem-solving and mastery paths all run on one agent loop instead of being six separate features glued to a UI.
The stack is unusually honest about being a harness rather than a model. It consults subagents — Claude Code, Codex, Gemini, Kimi — depending on the job. RAG is pluggable across LlamaIndex, PageIndex, GraphRAG and LightRAG rather than committing to one retrieval religion. MCP servers and community skills drop in as tools. FastAPI backend, Next.js 16 frontend, optional runner sandbox, single-user or multi-user with isolated workspaces. Apache 2.0.
The three-layer memory system with inspectable personalization is the part that separates this from a chatbot with a textbook attached. A tutor that doesn't remember what you got wrong last week is not a tutor. A tutor whose memory of you is a black box is worse. Letting the learner open up and read what the system thinks it knows about them is the right call, and almost nobody does it.
This is where the 100X thesis stops being about code. Agents that write software are the visible frontier, but agents that compress the time it takes a human to actually understand something are the bigger multiplier — you only write the code once, you carry the understanding into everything after. Version 1.5.11 landed this month with fixes to model output handling and memory tracking; releases are frequent.
https://github.com/HKUDS/DeepTutor
← Back to all articles
The stack is unusually honest about being a harness rather than a model. It consults subagents — Claude Code, Codex, Gemini, Kimi — depending on the job. RAG is pluggable across LlamaIndex, PageIndex, GraphRAG and LightRAG rather than committing to one retrieval religion. MCP servers and community skills drop in as tools. FastAPI backend, Next.js 16 frontend, optional runner sandbox, single-user or multi-user with isolated workspaces. Apache 2.0.
The three-layer memory system with inspectable personalization is the part that separates this from a chatbot with a textbook attached. A tutor that doesn't remember what you got wrong last week is not a tutor. A tutor whose memory of you is a black box is worse. Letting the learner open up and read what the system thinks it knows about them is the right call, and almost nobody does it.
This is where the 100X thesis stops being about code. Agents that write software are the visible frontier, but agents that compress the time it takes a human to actually understand something are the bigger multiplier — you only write the code once, you carry the understanding into everything after. Version 1.5.11 landed this month with fixes to model output handling and memory tracking; releases are frequent.
https://github.com/HKUDS/DeepTutor
Comments