DeepTutor is an agent-native personalized tutoring platform written in Python that implements lifelong learning through AI-powered multi-agent systems. The project is hosted at deeptutor.info and is documented in the research paper available at arxiv.org/abs/2604.26962. The repository is classified across multiple educational technology domains including intelligent tutoring, reinforcement learning, deep learning, personalized learning, adaptive learning, student modeling, and pedagogical strategies.
The platform operates as a comprehensive tutoring ecosystem built on large language models and multi-agent architectures. Core features include interactive learning capabilities, retrieval-augmented generation (RAG) functionality, deep research agents, and a command-line interface tool. The system supports multiple retrieval engines including GraphRAG, PageIndex, LightRAG, and linked knowledge bases, with document parsing capabilities that handle multimodal content including image extraction. Users can connect custom AI partners as agents that maintain their own personas, libraries, and skills with private memory isolation.
The architecture includes several specialized components. The Guided Learning system operates on a chat agent loop with mastery gates per question type and a dedicated learning dashboard. A Memory system operates across three layers (L1/L2/L3) for context management. The platform supports 15 different communication channels through its production-grade instant messaging pipeline, including native support for Mattermost, Matrix, Zulip, and other platforms. The TutorBot component provides tool sandboxing with per-user resource isolation and can be accessed through HTTP/SSE APIs.
Development activity shows strong engagement with 84 tracked issues and pull requests. The median response latency for issues and PRs is 0.0 hours with a mean of 21.4 hours, indicating active maintenance. Bug reports comprise the most active label category with 30 items, followed by enhancement requests with 22 items and questions with 14 items. The primary contributor pancacake has logged 169 events, with desic and yanOnGithub contributing 22 and 21 events respectively. The repository shares contributors with dbeaver/dbeaver and infiniflow/ragflow, indicating cross-project collaboration.
Recent releases demonstrate rapid iteration with multiple updates in June 2026. Version 1.5.0 introduced LlamaIndex document parsing with multimodal image extraction and Python 3.14+ compatibility. Version 1.4.15 added native Mattermost channel support and fixed grading in Guided Learning. Version 1.4.12 introduced the LightRAG Server retrieval engine and PyMuPDF4LLM parsing engine with FAISS vector backend support for faster large knowledge-base retrieval. Earlier versions established core features including native tool calling across OpenAI-compatible providers, self-service profile pages with avatars, and the ability to connect local Claude Code instances.
The platform supports flexible deployment options including containerized setups with rootless Podman compatibility and single-port request-time proxy configurations. Knowledge base functionality includes real in-browser previews for DOCX and XLSX files. The system implements security features including deny-by-default MCP tools for non-admin users and auth-routing mechanisms. Community engagement is facilitated through ClawHub integration, allowing users to install community skills via the deeptutor CLI with security gates. The project actively welcomes contributions through a documented contributing guide covering branching strategy and coding standards.
