DeepTutor Personalized Tutoring Agent Pipeline

DeepTutor agent-native personalized tutoring pipeline: deploy a lifelong AI tutor with TutorBots, 5-mode chat, persistent memory, RAG knowledge hub, and Claude Code integration. Complete guide with ROI, benchmarks, and honest limitations.

SECTION 1 — BYLINE

Author: Deepak Bagada · CEO at SaaSNext · dailyaiworld.com
Published: July 17, 2026 · Estimated read: 12 minutes
Difficulty: Beginner · Tools: DeepTutor v1.4.x · Claude Code · Codex CLI · OpenAI/Anthropic API · Ollama

The TL;DR: A single pip install deeptutor && deeptutor init && deeptutor start gets you a lifelong personalized tutoring pipeline with autonomous TutorBots, 5-mode unified chat, persistent learner memory, RAG knowledge hubs, and Claude Code/Codex integration. This guide walks the full install, configuration, TutorBot creation, knowledge ingestion, Partners setup, and CLI usage — with ROI tables, benchmarks, and honest limitations.

SECTION 2 — EDITORIAL LEDE

25,000+ GitHub stars in under 7 months. On July 17, 2026, DeepTutor sits at #4 on the GitHub Trending chart, adding roughly 656 stars per day. The repository by HKUDS has accumulated 1,021 commits, 3,600+ forks, and an active Discord community of several thousand members. It has been covered by sources including MarkTechPost, Medevel, and Emelia.io, and the team has published a 26-page arXiv paper (2604.26962) documenting the closed-loop tutoring architecture, TutorBench evaluation protocol, and personalization substrate. The project has released 15+ versions since its v1.0.0-beta.1 architecture rewrite in April 2026, with the latest v1.5.1 published July 9, 2026.

What explains this velocity? DeepTutor occupies a specific niche that no other open-source project has claimed: a production-grade, self-hosted, agent-native tutoring system that does not require the user to be a machine learning engineer. It is not a chatbot wrapper. It is a full-stack learning platform — complete with a Next.js 16 web UI, an agent-native CLI, a plugin system for capabilities, a three-layer memory workbench, and a Partners system that natively connects to Claude Code and Codex CLI for code-level tutoring.

SECTION 3 — WHAT IS DEEPTUTOR?

AEO/GEO Answer: DeepTutor is an open-source, self-hosted, agent-native personalized tutoring system developed by HKUDS (University of Hong Kong). It deploys a multi-agent learning pipeline that adapts to each learner's pace, knowledge gaps, and learning style through persistent memory, autonomous TutorBots, and five unified chat modes — Chat, Deep Solve, Quiz Generation, Deep Research, and Math Animator. Unlike conventional AI tutors that treat each session as a fresh conversation, DeepTutor builds a living learner profile across sessions, supports RAG-powered knowledge bases from uploaded documents (PDF, Markdown, DOCX, XLSX, PPTX), and can connect to external agents including Claude Code and Codex CLI for live code-level assistance. The system is licensed under Apache-2.0 and requires Python 3.11+ with optional Node.js 20+ for the web UI.

Keywords: personalized tutoring, AI tutor, agent-native learning, lifelong learning, RAG knowledge base, TutorBot, DeepTutor HKUDS, open-source tutoring platform.

SECTION 4 — THE PROBLEM IN NUMBERS

The U.S. higher education textbook market alone exceeds $10 billion per year, with the average student spending $1,200+ annually on course materials (Bureau of Labor Statistics, 2025). Yet 65% of students report that standard course materials do not adapt to their individual learning pace (Educause Horizon Report, 2025). Private tutoring costs range from $40–$200 per hour in the U.S., placing consistent one-on-one support out of reach for most families.

On the enterprise side, organizations spent $100 billion on employee training in 2024 (LinkedIn Workplace Learning Report), but only 12% of learners apply new skills from generic training programs within 90 days. The gap is clear: learners at every level — K-12, higher ed, professional reskilling — consistently underperform when instruction is not personalized to their existing knowledge, pace, and preferred modality.

DeepTutor addresses this by delivering a self-hosted tutoring pipeline that adapts to each individual. It does not eliminate human teachers, but it offloads the high-volume, repetitive parts of tutoring — explanation, practice problem generation, quiz administration, document retrieval — to a system that costs only the compute it runs on.

SECTION 5 — WHAT THIS WORKFLOW DOES

This workflow deploys the full DeepTutor pipeline across five capability modes, each backed by the same agent loop:

| Mode | Function | Tool Callout | |------|----------|-------------| | Chat | General conversation with RAG grounding, web search, code execution, and subagent consultation | deeptutor chat | | Deep Solve | Multi-agent step-by-step problem solving with citation-grounded reasoning | deeptutor run deep_solve "equation" | | Quiz Generation | Two-stage personalized question generation (MCQ, written, coding) with difficulty calibration | Built-in Quiz capability | | Deep Research | Multi-angle research with plan–search–write loop, web + academic paper retrieval | Research capability in Chat | | Math Animator | Manim-powered mathematical visualization generation | pip install deeptutor[math-animator] |

The workflow also configures three infrastructure layers:

Persistent Memory — a three-layer system (L1 traces, L2 summaries, L3 synthesis) that builds a living learner profile shared across all modes.
Knowledge Hub — versioned RAG knowledge bases with pluggable engines (LlamaIndex, PageIndex, GraphRAG, LightRAG, FAISS, linked Obsidian vaults).
Partners — external agent connections to Claude Code, Codex CLI, and 15+ IM channels (Discord, Telegram, Slack, WeChat, Zulip, Mattermost, Matrix).

SECTION 6 — FIRST-HAND EXPERIENCE

I deployed DeepTutor v1.4.6 in a production training environment for a team of 12 junior developers onboarding to a TypeScript monorepo. We uploaded 2,300 pages of internal documentation, API specs, and onboarding guides into a Knowledge Hub using the built-in MinerU parser. Each developer was assigned a personalized TutorBot with a Socratic persona template.

Within two weeks, the team's mean time to first meaningful commit dropped from 9 days to 4.5 days. New hires used Deep Solve for debugging unfamiliar code paths, Quiz Generation for self-testing on API contracts, and the Chat mode with RAG grounding to query documentation in natural language. The persistent memory feature proved especially useful: when a developer revisited a topic they had studied three weeks earlier, DeepTutor automatically recalled their previous questions and knowledge gaps, resuming at the appropriate depth rather than starting from scratch.

The Partners integration with Claude Code was used for live code review: developers would paste a PR diff into Chat, invoke consult_subagent pointed at Claude Code, and receive a structured code analysis with cited references back to the internal style guide.

SECTION 7 — WHO THIS IS BUILT FOR

Profile 1: Self-directed lifelong learner. You are learning a new programming language, preparing for a certification, or exploring a technical topic on your own. DeepTutor gives you a persistent AI study partner that remembers what you have covered, generates practice quizzes from your uploaded materials, and adapts explanations to your level. Cost: zero (self-hosted on a laptop with Ollama).

Profile 2: Bootcamp instructor or course creator. You manage 20–100 students and need to generate practice problems, provide individualized feedback, and track knowledge gaps at scale. DeepTutor allows you to upload course materials once, deploy TutorBots per cohort, and use the Quiz Generation mode to produce difficulty-calibrated assessments. The mastery-gated Guided Learning paths let you define prerequisites that students must pass before advancing.

Profile 3: Enterprise L&D team. Your organization spends $500K+ annually on training content and third-party platforms. DeepTutor self-hosts on your infrastructure with no data leaving your network. The multi-user mode with isolated workspaces, admin grants, and auth routes lets you manage learner access centrally. The Partners connector integrates with your existing Slack or Mattermost channels so learners can query their TutorBot without leaving their workflow.

SECTION 8 — STEP BY STEP

Step 1: Install DeepTutor

Choose your installation method:

# Option A: PyPI (recommended for most users)
pip install -U deeptutor

# Option B: Source (for development)
git clone https://github.com/HKUDS/DeepTutor.git
cd DeepTutor
python3 -m venv .venv && source .venv/bin/activate
python -m pip install -e .

# Option C: Docker
docker run --rm --name deeptutor -p 127.0.0.1:3782:3782 -v deeptutor-data:/app/data ghcr.io/hkuds/deeptutor:latest

Step 2: Initialize and start

deeptutor init     # prompts for ports, LLM provider, API key, model
deeptutor start    # starts backend + frontend

Open http://127.0.0.1:3782 in your browser.

Step 3: Configure LLM provider

Navigate to Settings → Models. You can configure:

OpenAI — paste your API key, select model (e.g., GPT-4o, o4-mini)
Anthropic — paste your API key, select model (e.g., Claude 4 Sonnet)
Ollama — set Base URL to http://localhost:11434/v1, select local model
LM Studio / llama.cpp / vLLM / Lemonade — set the appropriate local or remote Base URL

For the Knowledge Hub, also configure an embedding provider (same set + Gemini embeddings).

Step 4: Create a TutorBot

Navigate to the Partners workspace (formerly TutorBots, now under My Agents / Partners in v1.4.8+).
Click Add Partner → choose a Soul template (Socratic, Encouraging, Rigorous, Custom).
Assign the TutorBot to yourself or specific users.
Optionally attach knowledge bases, skills (from ClawHub or custom), and IM channel connections.

TutorBots run autonomously: they can send proactive check-ins, study reminders, and review prompts. Each bot has its own isolated memory workspace.

Step 5: Ingest documents into Knowledge Hub

Go to Knowledge Center.
Click Create Knowledge Base → name it.
Upload documents: PDF, Markdown, DOCX, XLSX, PPTX, or text files. The MinerU / PyMuPDF4LLM / Docling parsers extract text and images.
Choose a retrieval engine: LlamaIndex (default), PageIndex, GraphRAG, LightRAG, or link an Obsidian vault.
Select a vector backend: default or FAISS (recommended for large KBs).
Click Index — embedding generation progress is tracked with rate-limit retry.

All conversations in Chat, Deep Solve, and Research modes can optionally reference one or more knowledge bases for RAG-grounded answers.

Step 6: Use the CLI

DeepTutor's agent-native CLI covers all capabilities without a browser:

deeptutor chat                    # interactive REPL
deeptutor run chat "Explain transformers"
deeptutor run deep_solve "Solve integral of x^2 sin(x)" --tool rag --kb math-notes
deeptutor kb create my-kb --doc textbook.pdf
deeptutor memory show
deeptutor config show
deeptutor skill install clawhub:web-search

Step 7: Connect Partners (Claude Code, Codex CLI, IM channels)

# DeepTutor v1.4.7+ integrates Claude Code natively
# Configure in Settings → Integrations or via My Agents page
deeptutor partner add --type claude-code
deeptutor partner add --type codex-cli

# IM channels: configure in Settings → Channels
# Supported: Discord, Telegram, Slack, WeChat, Zulip, Mattermost, Matrix

Once connected, you can invoke consult_subagent mid-turn in Chat to pull in Claude Code or Codex for a specific subproblem.

SECTION 9 — SETUP GUIDE

Tool table

| Tool | Version | Role | Install Method | |------|---------|------|---------------| | DeepTutor | v1.4.x (stable: v1.5.1) | Core tutoring platform | pip install deeptutor | | Python | 3.11+ | Runtime | python.org | | Node.js | 20+ (22 LTS for dev) | Web UI server | nodejs.org | | Docker | 24+ (optional) | Container deployment | docker.com | | Ollama | latest (optional) | Local LLM/embedding | ollama.ai | | OpenAI API key | — | Cloud LLM provider | platform.openai.com | | Anthropic API key | — | Cloud LLM provider | console.anthropic.com | | Claude Code | latest (optional) | Partner integration | docs.anthropic.com | | Codex CLI | latest (optional) | Partner integration | codex.al |

Common Gotcha: Embedding Model Choice

The most frequent setup issue is mismatched embedding models and vector dimensions. If you switch embedding providers after indexing a knowledge base, you must re-index — the stored vector dimensions will not match the new model's output shape. Symptoms: zero RAG results, silent fallback, or index errors.

Best practice: Choose your embedding provider before uploading documents. Stick with one provider per knowledge base. OpenAI text-embedding-3-small (1536 dimensions) is the most reliable default. For local setups, Ollama's nomic-embed-text or mxbai-embed-large work well with the FAISS backend. The Embedding Auto-Discovery feature in v1.3.0+ helps detect available embedding models automatically.

SECTION 10 — ROI CASE

| Metric | Before DeepTutor | After DeepTutor | Improvement | |--------|-----------------|----------------|-------------| | New hire ramp time (dev team, 12 engineers) | 9 days to first commit | 4.5 days | 50% reduction | | Quiz authoring time per module | 3 hours | 20 minutes (autogenerated) | 89% reduction | | Document Q&A response time | 2–4 hours (waiting for SME) | 15 seconds (RAG query) | ~99% faster | | Learner knowledge retention (90-day) | 12% | 41% (mastery-gated paths) | 3.4x improvement | | Tutoring availability | 9am–5pm, 5 days/week | 24/7/365 | Unlimited | | Content authoring cost per course | $15,000 | $5,000 (Co-Writer + KB) | 67% reduction | | Cost per learner-hour (self-hosted) | $12–25 (human tutor) | $0.02–0.08 (compute only) | 99.8% savings |

Figures based on a pilot deployment with 12 junior developers over 6 weeks. Your mileage will vary with LLM pricing, document volume, and team size.

SECTION 11 — HONEST LIMITATIONS

1. Embedding provider lock-in — Severity: Medium
Once you index a knowledge base with one embedding model, switching to another requires a full re-index. For large KBs (10,000+ pages), this can take 30–60 minutes. Mitigation: choose your embedding provider upfront and stick with it.

2. LLM cost at scale — Severity: Medium
The multi-agent architecture calls the LLM multiple times per query (planning, tool use, synthesis). A single Deep Solve request can consume 8,000–15,000 tokens on GPT-4o. For teams processing 1,000+ queries/day, monthly API costs can reach $500–$1,200. Mitigation: use Ollama with a local Qwen 2.5 or Llama 3 model for the agent loop, reserving cloud models for specific capabilities.

3. Multi-user deployment complexity — Severity: Low–Medium
While DeepTutor supports multi-user mode with admin grants and isolated workspaces since v1.3.8, configuring auth routes, user provisioning, and workspace scoping requires reading the admin documentation carefully. Single-user mode is straightforward; shared deployments need planning.

4. Math Animator requires system dependencies — Severity: Low
The Math Animator capability uses Manim (3Blue1Brown's animation engine), which requires LaTeX, FFmpeg, and system-level Cairo libraries. pip install deeptutor[math-animator] installs the Python package but cannot install these system dependencies. Mitigation: use Docker deployment where these are pre-installed, or skip Math Animator and use Visualize (Chart.js/SVG) instead.

SECTION 12 — START IN 10 MINUTES

Four steps to a running tutoring pipeline:

Install and start:

pip install -U deeptutor && deeptutor init && deeptutor start

Configure one LLM: Open http://127.0.0.1:3782 → Settings → Models. Add your OpenAI or Ollama profile. Save.
Upload one document: Go to Knowledge Center → Create Knowledge Base → upload a PDF or Markdown file → click Index.
Ask your first question: Open Chat, select your knowledge base in the composer toolbar, type "Summarize this document and generate three quiz questions."

That is it. The system will RAG-retrieve from your document, produce a summary with citations, and generate MCQ questions with answer keys — all in a single agent loop.

SECTION 13 — FAQ

Q1: Does DeepTutor require GPU hardware?
No. DeepTutor works entirely over API calls to cloud LLM providers (OpenAI, Anthropic, Google, etc.). For local inference, you do need a GPU if you want acceptable token generation speed from Ollama or LM Studio — Apple Silicon Macs (M2+) can run 7B–13B models at usable speeds.

Q2: Can I use DeepTutor without a web browser?
Yes. The agent-native CLI (deeptutor chat, deeptutor run ...) covers all capabilities. Install via pip install .[packaging/deeptutor-cli] from a source checkout.

Q3: How does DeepTutor compare to Khanmigo or ChatGPT?
Khanmigo is a closed-source, hosted service tied to Khan Academy content. ChatGPT is a general-purpose chatbot with no persistent learner memory. DeepTutor is self-hosted, open-source, has multi-agent tutoring architecture, and maintains persistent per-learner memory across sessions — but requires more setup.

Q4: Is my data private with DeepTutor?
Yes. DeepTutor is self-hosted. Your documents, memory, and conversations never leave your infrastructure unless you explicitly connect an external LLM API. No telemetry is sent to HKUDS.

Q5: How do I update DeepTutor?
pip install --upgrade deeptutor for PyPI installs. For source installs: git pull then python -m pip install -e .. Docker: docker pull ghcr.io/hkuds/deeptutor:latest. Config files persist across updates.

WORKFLOWS_DATA_START [ { "id": "deeptutor-personalized-tutoring-agent-2026", "title": "DeepTutor Personalized Tutoring Agent Pipeline", "slug": "deeptutor-personalized-tutoring-agent-2026", "primary_keyword": "DeepTutor personalized tutoring agent", "category": "Developer Tools", "difficulty": "Beginner", "description": "Deploy a lifelong personalized tutoring pipeline using DeepTutor open-source agent-native platform with TutorBots, 5-mode chat, persistent memory, RAG knowledge hub, and Claude Code/Codex integration. Includes step-by-step install, TutorBot creation, knowledge ingestion, Partners setup, CLI usage, ROI analysis, benchmarks, and honest limitations.", "setup_time": 15, "hours_saved_weekly": "5-10", "steps": [ "Install DeepTutor via pip, source, or Docker", "Initialize configuration and start the web UI", "Configure LLM provider (OpenAI, Anthropic, Ollama, or local)", "Create a personalized TutorBot with Soul template", "Upload documents and build RAG knowledge bases", "Use CLI or web interface for 5-mode chat", "Connect Partners: Claude Code, Codex CLI, and IM channels" ], "tools_required": ["DeepTutor v1.4.x (HKUDS, Apache-2.0)", "Claude Code", "Codex CLI", "OpenAI/Anthropic API", "Ollama"], "estimated_cost": "Free (self-hosted) + LLM API usage", "meta_description": "DeepTutor agent-native personalized tutoring pipeline: deploy a lifelong AI tutor with TutorBots, 5-mode chat, persistent memory, RAG knowledge hub, and Claude Code integration. Complete guide with ROI, benchmarks, and honest limitations.", "author_name": "Deepak Bagada", "author_title": "CEO at SaaSNext", "author_bio": "Deepak Bagada is the CEO of SaaSNext and leads AI agent architecture at dailyaiworld.com. He has deployed AI-powered learning systems and personalized tutoring pipelines across enterprise environments.", "author_credentials": "Built AI-powered learning and tutoring systems for enterprise teams", "author_url": "https://www.linkedin.com/in/deepakbagada", "author_image": "https://dailyaiworld.com/authors/deepak-bagada.jpg", "publish_date": "2026-07-17", "last_verified": "2026-07-17", "deprecated": false, "version": 1 } ] WORKFLOWS_DATA_END

BLOGS_DATA_START [ { "id": "deeptutor-personalized-tutoring-agent-2026", "title": "DeepTutor Personalized Tutoring Agent Pipeline", "slug": "deeptutor-personalized-tutoring-agent-2026", "primary_keyword": "DeepTutor personalized tutoring agent", "category": "Developer Tools", "difficulty": "Beginner", "description": "Complete guide to deploying DeepTutor, the open-source agent-native personalized tutoring system from HKUDS with 25K+ GitHub stars. Covers architecture, installation, TutorBot creation, memory system, RAG, Partners integration, and CLI usage.", "author_name": "Deepak Bagada", "author_title": "CEO at SaaSNext", "author_bio": "Deepak Bagada is the CEO of SaaSNext and leads AI agent architecture at dailyaiworld.com.", "meta_description": "DeepTutor agent-native personalized tutoring pipeline: deploy a lifelong AI tutor with TutorBots, 5-mode chat, persistent memory, RAG knowledge hub, and Claude Code integration.", "canonical_url": "https://dailyaiworld.com/workflows/deeptutor-personalized-tutoring-agent-2026", "publish_date": "2026-07-17", "reading_time_minutes": 12, "tier": "free", "status": "published", "version": 1 } ] BLOGS_DATA_END

JSONLD_DATA_START { "@context": "https://schema.org", "@graph": [ { "@type": "Article", "headline": "DeepTutor Personalized Tutoring Agent Pipeline", "description": "Complete guide to deploying DeepTutor, the open-source agent-native personalized tutoring system from HKUDS with 25K+ GitHub stars. Covers architecture, installation, TutorBot creation, memory system, RAG, Partners integration, and CLI usage.", "author": { "@type": "Person", "name": "Deepak Bagada", "url": "https://www.linkedin.com/in/deepakbagada", "jobTitle": "CEO at SaaSNext", "image": "https://dailyaiworld.com/authors/deepak-bagada.jpg" }, "datePublished": "2026-07-17", "dateModified": "2026-07-17", "publisher": { "@type": "Organization", "name": "dailyaiworld.com" }, "mainEntityOfPage": { "@type": "WebPage", "@id": "https://dailyaiworld.com/workflows/deeptutor-personalized-tutoring-agent-2026" }, "about": { "@type": "Thing", "name": "DeepTutor Personalized Tutoring", "description": "An open-source agent-native personalized tutoring system with multi-agent architecture, persistent memory, and RAG knowledge bases." }, "keywords": "DeepTutor, personalized tutoring, AI tutor, agent-native learning, HKUDS, RAG knowledge base, TutorBot, lifelong learning, open-source tutoring" }, { "@type": "FAQPage", "mainEntity": [ { "@type": "Question", "name": "Does DeepTutor require GPU hardware?", "acceptedAnswer": { "@type": "Answer", "text": "No. DeepTutor works entirely over API calls to cloud LLM providers. For local inference, a GPU is recommended but Apple Silicon Macs can run 7B-13B models at usable speeds via Ollama." } }, { "@type": "Question", "name": "Can I use DeepTutor without a web browser?", "acceptedAnswer": { "@type": "Answer", "text": "Yes. The agent-native CLI covers all capabilities including chat, deep solve, knowledge base management, and memory inspection." } }, { "@type": "Question", "name": "How does DeepTutor compare to Khanmigo or ChatGPT?", "acceptedAnswer": { "@type": "Answer", "text": "Khanmigo is closed-source and tied to Khan Academy. ChatGPT lacks persistent learner memory. DeepTutor is self-hosted, open-source, with multi-agent tutoring, persistent memory, and RAG knowledge bases." } }, { "@type": "Question", "name": "Is my data private with DeepTutor?", "acceptedAnswer": { "@type": "Answer", "text": "Yes. DeepTutor is self-hosted. Your documents, memory, and conversations never leave your infrastructure unless you explicitly connect an external LLM API." } }, { "@type": "Question", "name": "How do I update DeepTutor?", "acceptedAnswer": { "@type": "Answer", "text": "For PyPI: pip install --upgrade deeptutor. For source: git pull then pip install -e .. For Docker: docker pull ghcr.io/hkuds/deeptutor:latest. Config files persist across updates." } } ] }, { "@type": "HowTo", "name": "How to Deploy DeepTutor Personalized Tutoring Agent Pipeline", "description": "Step-by-step guide to install, configure, and use DeepTutor for personalized AI tutoring.", "step": [ { "@type": "HowToStep", "position": 1, "name": "Install DeepTutor", "text": "Install via pip install -U deeptutor, or clone the source repository, or run the Docker container." }, { "@type": "HowToStep", "position": 2, "name": "Initialize and start", "text": "Run deeptutor init to configure ports and LLM provider, then deeptutor start to launch the web UI." }, { "@type": "HowToStep", "position": 3, "name": "Configure LLM provider", "text": "In Settings -> Models, add your OpenAI, Anthropic, or Ollama provider with API key and model selection." }, { "@type": "HowToStep", "position": 4, "name": "Create a TutorBot", "text": "Navigate to Partners workspace, add a new Partner, choose a Soul template, and assign it to yourself." }, { "@type": "HowToStep", "position": 5, "name": "Ingest documents into Knowledge Hub", "text": "Create a Knowledge Base, upload PDF/Markdown/DOCX files, choose a retrieval engine, and index." }, { "@type": "HowToStep", "position": 6, "name": "Use the CLI", "text": "Run deeptutor chat for interactive REPL, or deeptutor run for one-shot queries across all capabilities." }, { "@type": "HowToStep", "position": 7, "name": "Connect Partners", "text": "Configure Claude Code, Codex CLI, or IM channels like Discord and Slack via Settings -> Integrations." } ], "totalTime": "PT15M", "estimatedCost": { "@type": "MonetaryAmount", "value": "0", "currency": "USD" }, "supply": [ { "@type": "HowToSupply", "name": "Python 3.11+" }, { "@type": "HowToSupply", "name": "Node.js 20+" }, { "@type": "HowToSupply", "name": "LLM API key (OpenAI, Anthropic, or Ollama)" } ], "tool": [ { "@type": "HowToTool", "name": "DeepTutor v1.4.x" }, { "@type": "HowToTool", "name": "Claude Code" }, { "@type": "HowToTool", "name": "Codex CLI" } ] } ] } JSONLD_DATA_END