DeepTutor is an open-source AI tutoring workspace from the HKUDS group that you install and run yourself. Melo is explainx.ai's hosted learning copilot that you open in a browser. Both try to fix the same problem, chatbots that answer but do not teach, and they take opposite routes to get there.
DeepTutor crossed 50.5k GitHub stars and shipped v1.7.0 on October 28, 2026, so it is showing up in a lot of "what should I use to study" threads. This post explains what it is first, then compares it with Melo on the questions people actually ask: setup, cost, grounding, grading and who each one is for.
TL;DR
| Question | DeepTutor | explainx.ai Melo |
|---|---|---|
| What is it? | Open-source, self-hosted learning workspace | Hosted AI learning copilot inside explainx.ai |
| Setup | pip install, deeptutor init, or Docker; bring your own model keys | None. Open /dashboard/melo and start |
| License / cost | Apache-2.0; you pay for model APIs or hardware | Free plan; Monthly $29, Annual $300 for larger allowances |
| Content | Whatever you upload or point it at | Curated explainx.ai pathways, articles, skills, MCP, tools, agents and loops registries |
| Grading | Quizzes, mastery gates, question banks | Quiz scored 1 / 0.5 / 0 against a visible rubric, flashcards, explain-back |
| Best for | Self-hosters and teams indexing private material | People who want to learn AI fast without running infrastructure |
What is DeepTutor?
DeepTutor describes itself as "Lifelong Personalized Tutoring," an agent-native learning workspace that connects tutoring, problem solving, quiz generation, research, visualization and mastery practice. The backend is Python (3.11 to 3.14) with a Next.js 16 and React 19 frontend, and an arXiv preprint documents the design.
The project is large. The repository lists more than 3,100 commits on main, about 6.2k forks, and an Apache-2.0 license. Releases go back to v0.2.0 in January 2026, which tells you how fast it moves.
What is inside the workspace
Everything runs on one shared runtime, so a chat session can hand context to a quiz or a research report.
| Module | What it does |
|---|---|
| Chat | Default mode with tool calls, knowledge-base grounding, attachments and image generation |
| Guided Learning | Built on the chat agent loop with a per-type mastery gate and a learning dashboard |
| Learning Notebook | Formerly Co-Writer. Pages and blocks with an AI helper called Dito; edits are reviewed before you accept them |
| Research | Produces cited reports |
| Book | Compiles chosen sources into an interactive "living book" after you review a chapter outline |
| Knowledge Center | RAG libraries over several engines, including LlamaIndex (default), GraphRAG and LightRAG, with an eval command reporting Recall@k, nDCG and MRR |
| Memory | File-backed, three layers (traces, surface facts, synthesis) that you can read and edit |
| Partners | Persistent companions reachable from chat apps such as Telegram, Slack and Discord |
| Video Lesson / Audio Overview | Narrated lessons and spoken summaries |
It also speaks MCP, can consult coding agents such as Claude Code and Codex, and installs skills from a community hub. If you want a refresher on those building blocks, see our guides to MCP and agent skills.
How you run it
The documented paths are PyPI, Docker and a source checkout:
pip install -U "deeptutor[full]"
deeptutor init
deeptutor start
The app then serves on http://127.0.0.1:3782. A Docker image is published to GHCR, and a CLI-only mode exists for terminal use. Model support is broad: OpenAI-compatible endpoints, Anthropic, Gemini and others in the cloud, plus Ollama, LM Studio, llama.cpp, vLLM and Lemonade locally.
Self-hosted DeepTutor setup shown as a small house-shaped box with a glowing core standing for a local model on your own machine
Authentication is off by default, which means single-user. Turn it on and you get an admin workspace, isolated per-user workspaces and roles such as Standard and Learner, where Learner adds guardian oversight.
What it costs
The software is free. The maintainers say they have no paid online product, and quota errors are reported as-is rather than falling back to a paid provider. The real cost is yours: model API usage or GPUs, a machine to host it, and the time to keep it updated.
What people are asking about DeepTutor
Is it hard to set up? Not for a developer, and real work for everyone else. You need Python or Docker, a model key or a local runtime, and separate configuration for embeddings, search and speech. Some features add platform quirks: GraphRAG supports Python 3.11 to 3.13 only, native Windows needs FFmpeg for some speech adapters, and Docker users must point localhost at host.docker.internal to reach a model on the host.
Is it safe to let it run code? Code execution uses a runner sidecar or Linux bubblewrap, and falls back to a restricted subprocess that its own documentation calls best effort when stronger backends are missing. Treat that fallback as a convenience, not a sandbox. The same caution applies to any agent with a shell, which is why we wrote about sandboxing agent skills.
Is it stable? It is pre-1.x in places and changes quickly. Many releases are labeled fix releases, and setup guides point at the dev branch while the default branch is main. That is normal for a fast open-source project, but it means upgrades deserve a look at the release notes.
Does it prove that learners do better? No tutoring tool can claim that from a star count. The honest evidence base is the research: a Dartmouth AI tutor study with a measured effect size, and a Berkeley and CMU study finding that AI assistance reduced persistence in 10 minutes. Design matters more than model size, a point we expand in how AI tutors teach by making.
DeepTutor vs Melo: the comparison
What Melo is, briefly
Melo is explainx.ai's learning copilot, in beta and free to start. It has eight modes: Chat, Teach, ELI5, Quiz, Practice, Explain Back, Flashcards and Interview. It lives where you learn: the homepage, the full workspace at /dashboard/melo, inside every learning pathway, on skill pages through "Try now with Melo," and in workshops. The full tour is in Introducing Melo.
Side by side
| Dimension | DeepTutor | Melo |
|---|---|---|
| Time to first lesson | Install, configure models and embeddings, then start | Open the page and ask |
| Grounding | Your documents through RAG; quality depends on what you load and how you configure it | A curated library: explainx.ai articles, pathways and the skills, MCP, tools, agents and loops registries, with cited sources |
| Pedagogy | Many tools in one workspace; the learner chooses the path | Modes built for recall: Quiz, Practice and Explain Back make you produce the answer |
| Grading | Quiz generation, mastery gates, question banks | Per-question scoring (correct 1, partial 0.5, not yet 0) with a visible rubric and equivalent wording accepted |
| Interactive visuals | Visualize module | Generative UI in the teach flow and /interact simulations generated on demand |
| Model choice | Any provider, including local | Managed for you |
| Privacy model | Data stays on your machine if you self-host | Hosted on explainx.ai; attach up to three files per conversation |
| Community and humans | Discord, Feishu and WeChat communities | Social learning, workshops and live instructors on the same platform |
| Maintenance | You update, secure and back it up | Nothing to maintain |
| Cost | Free software plus model and hardware costs | Free plan; Monthly $29 or Annual $300 for bigger allowances |
Two matching comparison sheets side by side with one green tick row, showing DeepTutor and Melo evaluated on the same criteria
Why Melo is the better pick for learning AI
Four reasons, in order of how much they change your outcome.
- The content is already right. A tutor is only as good as what it teaches from. A self-hosted workspace starts empty and inherits whatever you feed it, plus the retrieval settings you chose. Melo answers from pathways and registries that the explainx.ai team maintains, so the explanation about MCP or agent skills points at material you can read and check.
- It is built to make you recall, not to impress you. Fluent answers create the feeling of understanding. Melo's Quiz, Practice and Explain Back modes withhold the answer until you commit to one, then grade it against a rubric. That mechanism is the point of the product.
- Zero setup means you actually start. The biggest failure mode for self-hosted learning tools is the weekend that goes into configuration and never reaches a first lesson. Melo costs you a click.
- It sits inside a learning loop. Pathways, the skills registry, workshops and live instructors share one account, so practice with Melo connects to a structured curriculum and to real people when you get stuck.
Flashcard mid-flip with a green dot on its back, representing Melo flashcards and recall practice
Where DeepTutor wins
Being fair helps you choose. Pick DeepTutor when:
- You must self-host for compliance, privacy or air-gapped work.
- You want local models and full control over which provider answers.
- You are indexing a large private corpus, such as internal manuals or a research library, and want to tune RAG engines and evaluate retrieval.
- You are a developer who enjoys extending the stack, adding skills, MCP services or messaging-app companions.
Melo's limits are the mirror image. It is hosted, so you cannot run it on your own server or swap the model. It is in beta. Usage is metered over a rolling five-hour window and a weekly window, and larger allowances need a paid plan.
A quick decision guide
| Your situation | Use |
|---|---|
| New to AI, want a path from basics to agents | Melo |
| Preparing for an interview or workshop on a deadline | Melo |
| Learning from explainx.ai skills, MCP servers or tools | Melo |
| Need everything on your own hardware | DeepTutor |
| Studying a private document set with custom retrieval | DeepTutor |
| Want to experiment with the tutoring stack itself | DeepTutor |
Try it today
To compare for yourself in ten minutes, open Melo, ask it to quiz you on a topic from a pathway, and answer in your own words. Then use Explain Back on the concept you scored lowest on. If you also run DeepTutor, load the same topic there and compare which one made you work harder to answer.
A prompt to start with:
Quiz me on MCP, five questions, one at a time. After each answer, score me and show the rubric. Then ask me to explain what I got wrong back to you in my own words.
If you want guided practice with people, the live workshops use the same Melo copilot with the cohort's resources attached. Related research on whether machine tutors can model learners is in our coverage of Microsoft's StudentSim, and another open-source take on learning agents is covered in LearnHarness.
Related reading
- Introducing Melo, explainx.ai's learning copilot
- AI tutors that teach by making: Khanmigo and Melo
- Melo generative UI for real-time learning
- Melo
/interact: live sandboxed simulations - Social learning with Melo
- Interactive AI learning pathways
- Dartmouth AI tutor study
- LearnHarness open-source learning harness
Repository details such as star counts, version numbers and feature lists are accurate as of October 30, 2026, and the project changes quickly. Check the project's own documentation before relying on a specific setting. Melo pricing and allowances are as published on explainx.ai at the same date.