---

🌎 The Vision

Traditional education scales poorly. Static curricula fail to adapt, and single-prompt LLMs lack the memory, pedagogical structure, and safety required for true learning.

EduPilot solves this by leveraging a decentralized Multi-Agent architecture. Instead of relying on a single omniscient LLM, we use LangGraph to act as a routing supervisor state-machine. It dynamically intercepts natural language intents and routes tasks to highly specialized, goal-oriented CrewAI expert agents. The result? A fully autonomous digital tutor that maintains permanent state, consults real textbooks via Vector RAG, and serves everything seamlessly over the newly minted Model Context Protocol (MCP).

---

⚡ The Tech Stack

We don't do monolithic architectures here. This is a modular, event-driven orchestration stack:

• Orchestration Layer: LangGraph (Supervisor State Machine)
• Agent Intelligence: CrewAI & LangChain (Lesson Planner, Doubt Resolver, Quiz Generator)
• Knowledge Retrieval & RAG: ChromaDB (Local Embeddings for ultra-low latency contextual retrieval)
• State & Memory Persistence: SQLite (Native Offline Storage for longitudinal mastery mapping)
• API & Extensibility: FastAPI (REST endpoints) + FastMCP (Claude Desktop Integration)
• Frontend App: Zero-dependency Vanilla HTML/CSS/JS (Glassmorphism Dark Mode)

---

🧠 The Agent Force

Microservice Agent	Primary Responsibility	Associated Capabilities
👑 The Orchestrator	Traffic Controller: Ingests the task limitlessly, classifies the semantic intent, fetches SQLite mastery memory, and routes to the correct Crew.	Routing, State Augmentation, Guardrails
🧑‍🏫 Lesson Personalizer	Dynamic Curriculum: Composes 5E-Model tailored lesson plans dynamically adjusted for the student's exact learning style (Visual, Auditory, Kinesthetic) and age.	Bloom's Taxonomy Scaling, Adaptive Difficulty
🛡️ Doubt Resolver	RAG Explainer: Triggers a Vector DB retrieval across curriculum textbooks to answer questions accurately without hallucinating non-school-board facts.	ChromaDB, Pinecone, Misconception Bridging
📝 Quiz Generator	Assessment: Generates precise, misconception-targeted distractors for MCQs.	Formative Assessment
📊 Progress Tracker	Memory Core: Parses session outputs and permanently updates student mastery levels (0-100) inside the persistent SQLite memory layer.	Spaced Repetition Data, Database Hydration

---

📦 Quick Start Installation

Get up and running in your local dev environment in under 60 seconds.

1. Clone & Set up the Virtual Environment

python -m venv venv
source venv/bin/activate  # On Windows: .\venv\Scripts\Activate.ps1

2. Hydrate Dependencies

pip install -r requirements.txt

3. Inject Environment Keys Copy .env.example to .env and configure your foundation model provider (Anthropic is recommended for reasoning, OpenAI for generation):

ANTHROPIC_API_KEY="sk-ant-..."
OPENAI_API_KEY="sk-proj-..."

---

🎯 Running The Platform

EduPilot is designed to run anywhere. Choose your preferred interaction method:

Method A: The Interactive Web App (Recommended)

Boot the REST backend:

uvicorn main:app --reload --host 0.0.0.0 --port 8000

Then simply double-click frontend/index.html on your local machine to launch our completely decoupled, zero-build-step, ultra-premium chat interface.

Method B: Native MCP Integration (Claude Desktop)

Because EduPilot is an official MCP Server, you can pipe the agent orchestrator directly into Claude!

Simply modify your claude_desktop_config.json file:

{
  "mcpServers": {
    "edu-pilot-agent": {
      "command": "C:/path/to/venv/Scripts/python.exe",
      "args": ["C:/path/to/mcp_server.py"]
    }
  }
}

Note: Make sure to point to the absolute path of the python.exe inside your virtual environment so dependencies resolve properly!

Method C: The MCP Inspector

Want to hit the bare-metal MCP tools locally?

npx @modelcontextprotocol/inspector .\venv\Scripts\python.exe mcp_server.py

---

Built with ❤️ for the future of AGI-driven education.