// the find
HKUDS/DeepTutor
DeepTutor: Agent-native Personalized Tutoring. https://deeptutor.info/.
DeepTutor is a self-hosted AI learning workspace — Python backend, Next.js frontend — that wraps an agent loop around tutoring, RAG-backed knowledge bases, persistent companions (Partners), quiz generation, research reports, and a 'living book' compiler. It targets students and educators who want a private, configurable alternative to hosted AI tutoring services. Daily release cadence since April 2026, 25k stars in about six months.
The subagent integration is more than a gimmick: it actually spawns Claude Code or Codex CLI processes and streams their work into the conversation via a consult_subagent tool, rather than just pasting transcripts. The RAG layer offers a real choice of retrieval engines — LlamaIndex, GraphRAG, LightRAG, PageIndex, Obsidian vault — and versioned indexes that keep the previous working index intact during a rebuild, which is something most RAG apps get wrong. The three-layer memory (L1 event trace, L2 surface summaries, L3 synthesis) is file-backed and inspectable rather than an opaque vector store you can't audit or correct. Installation is cleaner than most projects this complex: pip install deeptutor, deeptutor init, deeptutor start gets you a running full-stack app with a bundled Next.js server.
The default execution sandbox runs model-generated Python directly in the host process (sandbox_allow_subprocess=true), and the community skills feature from EduHub installs and runs third-party SKILL.md playbooks — two separate surfaces where a bad prompt or malicious skill author can execute code on your machine. The scope is genuinely unwieldy: 14+ distinct surfaces (Chat, Book, Partners, Co-Writer, My Agents, Mastery Path, Memory, Knowledge Center, etc.) with interdependent state, and the release notes already show regressions getting patched the next day. The Manim animation feature requires LaTeX, ffmpeg, and system graphics libs as hard dependencies, which will break silently on most server environments. This is an academic lab project (HKUDS at HK University); the commit graph shows bursts of activity that correlate with paper deadlines, and long-term maintenance after the lab moves on is an open question.