AI Book Agent MCP Server
An MCP (Model Context Protocol) server that provides AI assistants with intelligent access to ML textbook content for creating accurate, source-grounded documentation. This pure Python implementation uses local models for complete privacy and cost efficiency.
Overview
This project transforms authoritative ML textbooks into a knowledge service that any MCP-compatible AI assistant can access. Using local LLMs (Qwen) and embeddings (sentence-transformers), it creates a private, cost-effective RAG system exposed via the official Python MCP SDK.
Why MCP?
Traditional Approach Limitations
- Knowledge locked in a single application
- Users must switch between tools
- Cannot leverage AI assistants they already use
- Difficult to integrate with existing workflows
MCP Server Benefits
- Universal Access: Works with Claude Desktop, VS Code, and any MCP client
- Workflow Integration: Use book knowledge directly in your IDE or chat
- Composability: Combine with other MCP servers (filesystem, GitHub, etc.)
- Future-Proof: As MCP ecosystem grows, your book agent automatically works with new tools
Architecture
┌─────────────────────┐ ┌─────────────────────┐
│ Claude Desktop │ │ VS Code │
│ "Explain ML drift │ │ "Generate docs │
│ from textbooks" │ │ for monitoring" │
└──────────┬──────────┘ └──────────┬──────────┘
│ │
└─────────────┬─────────────┘
│ MCP Protocol (stdio/HTTP)
▼
┌─────────────────────────────────────┐
│ Python MCP Server │
│ (Official modelcontextprotocol/ │
│ python-sdk) │
├─────────────────────────────────────┤
│ MCP Tools: │
│ - @mcp.tool() search_books() │
│ - @mcp.tool() get_chapter_content() │
│ - @mcp.tool() generate_section() │
│ - @mcp.tool() cite_sources() │
├─────────────────────────────────────┤
│ RAG Components (same process): │
│ - EPUB Parser (ebooklib) │
│ - Embeddings (sentence-transformers)│
│ - Vector Store (ChromaDB/FAISS) │
│ - LLM Generation (Ollama/Qwen) │
└─────────────────────────────────────┘
Technology Stack
- MCP Framework: Official Python SDK (
mcp[cli]) - Language: Python 3.11+
- Embeddings: sentence-transformers (all-MiniLM-L6-v2)
- LLM: Ollama with Qwen 2.5 7B
- Vector Store: ChromaDB/FAISS
- Book Parsing: ebooklib, BeautifulSoup4
- Transport: stdio (local) or HTTP (remote)
1. searchBooks
Search across all indexed books for relevant content
{
name: "searchBooks",
description: "Search ML textbooks for specific topics or concepts",
inputSchema: {
query: string,
bookFilter?: string[],
maxResults?: number,
includeContext?: boolean
}
}
2. getChapterContent
Retrieve specific chapter or section content
{
name: "getChapterContent",
description: "Get full content of a specific book chapter",
inputSchema: {
bookId: string,
chapterId: string,
format?: "markdown" | "plain"
}
}
3. generateSection
Generate documentation based on book content
{
name: "generateSection",
description: "Generate documentation section grounded in textbook sources",
inputSchema: {
topic: string,
outline?: string[],
style?: "technical" | "tutorial" | "overview",
maxSources?: number
}
}
4. citeSources
Get proper citations for content
{
name: "citeSources",
description: "Generate proper citations for book content",
inputSchema: {
bookId: string,
pageNumbers?: number[],
format?: "APA" | "MLA" | "Chicago"
}
}
Resources
The server exposes book content as browsable resources:
/books
├── /designing-ml-systems
│ ├── metadata.json
│ ├── /chapters
│ │ ├── /1-introduction
│ │ ├── /2-ml-systems-design
│ │ └── ...
│ └── /topics
│ ├── /monitoring
│ ├── /deployment
│ └── ...
└── /other-ml-book
└── ...
Prompts
Pre-configured prompts for common tasks:
doc_generator
name: doc_generator
description: Generate technical documentation from book sources
arguments:
- name: topic
description: The topic to document
- name: audience
description: Target audience (beginner/intermediate/advanced)
- name: length
description: Desired length (brief/standard/comprehensive)
concept_explainer
name: concept_explainer
description: Explain ML concepts using textbook definitions
arguments:
- name: concept
description: The concept to explain
- name: include_examples
description: Whether to include practical examples
Use Cases
1. In Claude Desktop
User: "Explain model drift using the ML textbooks"
Claude: [Uses searchBooks tool to find drift content]
[Retrieves relevant chapters]
[Generates explanation with citations]
2. In VS Code
# User comment: "TODO: Add monitoring based on best practices"
# AI Assistant uses book agent to generate monitoring code
3. Documentation Pipeline
# Automated doc generation using MCP tools
mcp-client generate-docs \
--server ai-book-agent \
--topics "deployment,monitoring,testing" \
--output ml-best-practices.md
Getting Started
Prerequisites
- Linux system (Ubuntu 22.04+ recommended)
- Python 3.11+
- Node.js 18+
- 16GB RAM minimum
- 20GB free disk space
Installation
# Clone the repository
git clone <repository-url>
cd ai-book-agent
# Install Python dependencies (with MCP SDK)
pip install "mcp[cli]" sentence-transformers chromadb ollama ebooklib beautifulsoup4
# Or use requirements.txt
pip install -r requirements.txt
# Install Ollama for local LLM
curl -fsSL https://ollama.ai/install.sh | sh
# Pull required models
ollama pull qwen2.5:7b
# Index existing books
python scripts/index_books.py
Configuration
Configure the server:
# config.yaml
embeddings:
model: "all-MiniLM-L6-v2"
device: "cpu" # or "cuda"
generation:
provider: "ollama"
model: "qwen2.5:7b"
base_url: "http://localhost:11434"
books:
data_dir: "data/epub"
index_dir: "data/vector_db"
Add to Claude Desktop config:
{
"mcpServers": {
"ai-book-agent": {
"command": "python",
"args": ["/path/to/ai-book-agent/server.py"]
}
}
}
Basic Usage
Once configured, the tools are available in any MCP client:
You: What does the <author> say about feature engineering?
Assistant: I'll search the ML textbooks for information about feature engineering.
[Calling searchBooks with query="feature engineering"]
[Found 5 relevant sections in "<Book Title>"]
According to <author> in "<Book Title>":
Feature engineering is described as... [content with citations]
Development
Adding New Books
# Place EPUB in data directory
cp new-ml-book.epub data/epub/
# Re-index all books
python scripts/index_books.py
# Server will automatically pick up new content
Add new tools directly in server.py:
from mcp.server.fastmcp import FastMCP
mcp = FastMCP("ai-book-agent")
@mcp.tool()
def compare_approaches(approach1: str, approach2: str) -> str:
"""Compare different ML approaches from multiple books"""
results1 = search_books(approach1, 3)
results2 = search_books(approach2, 3)
comparison = generate_comparison(results1, results2)
return comparison
@mcp.resource("book://{book_id}/summary")
def get_book_summary(book_id: str) -> str:
"""Get a summary of a specific book"""
return load_book_summary(book_id)
Testing
# Test MCP server locally
mcp dev server.py
# Test individual components
python scripts/test_components.py
python scripts/test_search.py
# Run full test suite
pytest tests/
# Test with Claude Desktop
mcp install server.py
Project Structure
ai-book-agent/
├── server.py # Main MCP server (entry point)
├── src/ # Python modules
│ ├── parsers/ # EPUB parsing
│ ├── embeddings/ # Embedding generation
│ ├── search/ # Vector search & retrieval
│ ├── generation/ # LLM integration (Ollama)
│ └── utils/ # Configuration and helpers
├── scripts/ # Utility scripts
│ ├── index_books.py # Index EPUB files
│ ├── test_*.py # Test individual components
│ └── setup.py # Initial setup
├── data/
│ ├── epub/ # Source EPUB files
│ ├── processed/ # Processed book content
│ └── vector_db/ # Vector store data
├── tests/ # Test suites
├── config.yaml # Configuration
├── requirements.txt # Python dependencies
└── README.md
Pure Python Architecture
This project uses a simplified, single-language approach:
- Python MCP Server: Official SDK handles MCP protocol and tool exposure
- Integrated RAG: All ML components run in the same Python process
- Local Models: Complete privacy with Ollama and sentence-transformers
Benefits of this approach:
- Simpler deployment: Single Python service
- Direct ML access: No API overhead between MCP and RAG
- Easier debugging: One codebase, one process
- Better performance: No network calls between components
Remote Access
For accessing the server from anywhere:
# Option 1: Cloudflare Tunnel (recommended)
cloudflared tunnel create book-agent
cloudflared tunnel run --url http://localhost:8080 book-agent
# Option 2: Configure with your domain
# See USER_GUIDE.md for detailed setup
Roadmap
Contributing
See USER_GUIDE.md for details on:
- Adding new tools
- Improving search algorithms
- Supporting new book formats
- Performance optimizations