CanvasXpress MCP Server

🎨 Model Context Protocol (MCP) server for generating CanvasXpress visualizations from natural language.

This MCP server provides AI assistants like Claude Desktop with the ability to generate CanvasXpress JSON configurations from natural language descriptions. It uses Retrieval Augmented Generation (RAG) with 132 few-shot examples (66 human + 66 GPT-4 descriptions) and semantic search.

📄 Based on: Smith & Neuhaus (2024) - CanvasXpress NLP Research
🔌 Supports: Azure OpenAI (BMS Proxy) or Google Gemini
🚀 Built with - The modern standard for Python MCP servers

🚀 Quick Start (5 Minutes)

Prerequisites

• ✅ Docker installed and running (or Python 3.10+ for local venv)

• ✅ API key: BMS Azure OpenAI or Google Gemini

• ✅ 8GB RAM for local embeddings, or 2GB for cloud embeddings

• ✅ No GPU required

Step 1: Clone Repository

git clone https://github.com/bms-ips/canvasxpress-mcp-server.git
cd canvasxpress-mcp-server

Step 2: Configure Environment

# Copy the example environment file
cp .env.example .env

# Edit .env with your credentials
nano .env  # or vim, code, etc.

For Azure OpenAI (BMS) with local embeddings:

LLM_PROVIDER=openai
AZURE_OPENAI_KEY=your_key_from_genai.web.bms.com
LLM_MODEL=gpt-4o-global
LLM_ENVIRONMENT=nonprod
EMBEDDING_PROVIDER=local  # BGE-M3 (highest accuracy)

For Google Gemini (fully cloud-based) ⭐ Lightweight:

LLM_PROVIDER=gemini
GOOGLE_API_KEY=your_key_from_aistudio.google.com
GEMINI_MODEL=gemini-2.0-flash-exp
EMBEDDING_PROVIDER=gemini  # Cloud embeddings (no PyTorch needed)

Step 3: Choose Your Setup

Option A: Docker (Full) - Default

make build      # Build Docker image (~8GB with PyTorch)
make init       # Initialize vector DB (downloads BGE-M3 ~2GB)
make run-http   # Start server

Option B: Local Venv (Lightweight) ⭐ For smaller servers

make venv-light  # Create lightweight venv (~500MB, no PyTorch)
make init-local  # Initialize vector DB (uses cloud embeddings)
make run-local   # Start server

Note: Option B requires EMBEDDING_PROVIDER=gemini or EMBEDDING_PROVIDER=openai in .env

Step 4: Test the Server

# Using the CLI client (works with either setup)
python3 mcp_cli.py -q "Create a bar chart with blue bars"

# Or inside Docker:
docker exec -it $(docker ps -q) python3 /app/mcp_cli.py -q "Create a scatter plot"

# Full JSON response
python3 mcp_cli.py -q "Bar chart" --json

That's it! 🎉

Your server is running at http://localhost:8000/mcp

Quick Reference:

📊 Server Modes

HTTP Mode (Network Access) - Recommended

make run-http   # Docker: daemon mode
make run-httpi  # Docker: interactive mode  
make run-local  # Local venv: foreground

• Access at: http://localhost:8000/mcp

• Use for: Remote access, web-based AI assistants, multiple clients

STDIO Mode (Local Only)

make run        # Docker
make run-locali # Local venv

• Use for: Claude Desktop integration on same machine

• Single client only

🧪 Testing Options

CLI Client (Quick & Easy) ⭐

python3 mcp_cli.py -q "Heatmap with clustering" --headers "Gene,Sample1,Sample2"
python3 mcp_cli.py -q "Bar chart" --json          # Full JSON response
python3 mcp_cli.py -q "Line chart" --config-only  # Config only

HTTP MCP Client

docker exec -it $(docker ps -q) python3 /app/mcp_http_client.py

Vector Database Test

make test-db  # Verifies 132 examples loaded

Python API

python3 examples_usage.py
# Tests the generator as a Python library (no MCP)

Next Steps:

• 🌐 HTTP Client Guide: See HTTP_CLIENT_GUIDE.md for HTTP/network usage

• 📖 Python API: See PYTHON_USAGE.md for API reference

• 🔌 Claude Desktop: Configure MCP integration (see below)

• 📘 Technical Details: See TECHNICAL_OVERVIEW.md for architecture

🌟 Features

• 🎯 High Accuracy: 93% exact match, 98% similarity (peer-reviewed methodology)

• 🔍 Semantic Search: BGE-M3 embeddings with vector database (or cloud embeddings)

• 🤖 Multi-Provider LLM: Azure OpenAI (BMS) or Google Gemini

• 📊 Multi-Provider Embeddings: Local BGE-M3, Azure OpenAI, or Google Gemini

• 🐳 Docker or Local: Run in Docker containers or Python virtual environment

• 🔌 FastMCP 2.0: Modern, Pythonic MCP server framework with HTTP & STDIO support

• 🌐 Network Access: HTTP mode for remote deployment and multiple concurrent clients

• 📊 132 Few-Shot Examples: 66 chart types × 2 description styles (human + GPT-4)

⚙️ Provider Configuration

The server supports multiple LLM and embedding providers. Configure in .env:

LLM Providers

Embedding Providers

ONNX Model Options

Set ONNX_EMBEDDING_MODEL to use different models:

Example Configurations

Azure OpenAI + Local BGE-M3 (highest accuracy, requires ~8GB RAM):

LLM_PROVIDER=openai
AZURE_OPENAI_KEY=your_key
LLM_MODEL=gpt-4o-global
LLM_ENVIRONMENT=nonprod
EMBEDDING_PROVIDER=local

Gemini + ONNX (lightweight local embeddings, ~500MB RAM) ⭐ Recommended for small servers:

LLM_PROVIDER=gemini
GOOGLE_API_KEY=your_key
GEMINI_MODEL=gemini-2.0-flash-exp
EMBEDDING_PROVIDER=onnx
ONNX_EMBEDDING_MODEL=all-MiniLM-L6-v2

Google Gemini + Local BGE-M3 (best accuracy with Gemini):

LLM_PROVIDER=gemini
GOOGLE_API_KEY=your_key
GEMINI_MODEL=gemini-2.0-flash-exp
EMBEDDING_PROVIDER=local

Full Gemini (no local model needed, faster startup):

LLM_PROVIDER=gemini
GOOGLE_API_KEY=your_key
GEMINI_MODEL=gemini-2.0-flash-exp
EMBEDDING_PROVIDER=gemini

Note: If you change EMBEDDING_PROVIDER, you must reinitialize the vector database (make init or make init-local) since different providers have different embedding dimensions.

📋 Prerequisites

• Docker OR Python 3.10+

• API key: BMS Azure OpenAI or Google Gemini

• Linux/macOS (tested on Amazon Linux 2023)

• 8GB RAM for local embeddings, or 2GB for cloud embeddings

• No GPU required (embeddings run on CPU)

🐍 Local Development (Alternative to Docker)

If you prefer running without Docker, you can use a Python virtual environment.

Prerequisites

⚠️ Requires Python 3.10+ (FastMCP requirement)

# Check your Python version
python3 --version

# If you have multiple Python versions, check for 3.10+
python3.10 --version  # or python3.11, python3.12

If Python 3.10+ is not installed, install it:

# Amazon Linux 2023 / RHEL 9 / Fedora
sudo dnf install python3.11 python3.11-pip python3.11-devel

# Amazon Linux 2 / RHEL 8 / CentOS 8
sudo amazon-linux-extras install python3.11
# Or use pyenv (see below)

# Ubuntu 22.04+
sudo apt install python3.11 python3.11-venv python3.11-dev

# Ubuntu 20.04 (add deadsnakes PPA first)
sudo add-apt-repository ppa:deadsnakes/ppa
sudo apt update
sudo apt install python3.11 python3.11-venv python3.11-dev

# macOS (with Homebrew)
brew install python@3.11

# Using pyenv (any Linux/macOS)
curl https://pyenv.run | bash
pyenv install 3.11.9
pyenv local 3.11.9

Step 1: Configure Python Path (if needed)

The Makefile uses python3.12 by default. If your Python 3.10+ has a different name:

# Check what Python executables you have
ls -la /usr/bin/python3*

# Edit the Makefile to use your Python
nano Makefile

# Change this line near the top:
PYTHON_BIN = python3.12
# To whatever you have, e.g.:
PYTHON_BIN = python3.11
# Or:
PYTHON_BIN = python3

Step 2: Create Virtual Environment

Option A: Full Installation (~8GB) - Local BGE-M3 embeddings

make venv

This creates a ./venv/ directory and installs all dependencies (~5-10 minutes for PyTorch/BGE-M3). Use this if you want EMBEDDING_PROVIDER=local (default, highest accuracy).

Option B: Lightweight Installation (~500MB) - Cloud embeddings only ⭐

make venv-light

This installs only cloud-compatible dependencies (no PyTorch, no BGE-M3). Perfect for lightweight servers that will use Gemini or OpenAI for embeddings.

⚠️ If using venv-light: You MUST set EMBEDDING_PROVIDER=gemini (or openai) in your .env file before running make init-local. The local BGE-M3 model is not installed.

Step 3: Configure Environment

cp .env.example .env
nano .env  # Add your AZURE_OPENAI_KEY

For lightweight venv, use this configuration:

# Example .env for venv-light (Gemini for both LLM and embeddings)
LLM_PROVIDER=gemini
GOOGLE_API_KEY=your_key_here
GEMINI_MODEL=gemini-2.0-flash-exp
EMBEDDING_PROVIDER=gemini
GEMINI_EMBEDDING_MODEL=text-embedding-004

Step 4: Initialize Vector Database

make init-local

Creates ./vector_db/canvasxpress_mcp.db with 132 embedded examples.

Step 5: Start Server

make run-local   # HTTP mode (http://localhost:8000/mcp)
make run-locali  # STDIO mode (for Claude Desktop)

Step 6: Test

# Activate venv for CLI
source venv/bin/activate
python mcp_cli.py -q "Create a bar chart"

Cleanup

make clean-local  # Remove venv and vector_db

Troubleshooting Local Setup

"python3.12: command not found"

• Edit PYTHON_BIN in the Makefile to match your Python 3.10+ executable name

• Common alternatives: python3.11, python3.10, python3

"No module named 'venv'" or venv creation fails

• Install the venv module: sudo apt install python3.11-venv (Ubuntu) or sudo dnf install python3.11 (RHEL/Amazon Linux)

Permission denied on vector_db

• If you previously ran Docker, the vector_db/ directory may be owned by root

• Fix with: sudo rm -rf vector_db && mkdir vector_db

Import errors when running

• Make sure you activated the venv: source venv/bin/activate

• Verify dependencies installed: pip list | grep fastmcp

"No module named 'FlagEmbedding'" (with venv-light)

• You used make venv-light but have EMBEDDING_PROVIDER=local in .env

• Fix: Set EMBEDDING_PROVIDER=gemini (or openai) in .env

• The lightweight venv doesn't include the BGE-M3 model

BGE-M3 model download fails

• Ensure you have ~2GB free disk space

• Check network connectivity (may need VPN for some networks)

• The model downloads to ~/.cache/huggingface/

🔧 Detailed Setup (Docker)

If you need more control or want to customize the Docker setup:

Build Docker Image

make build
# Or manually:
docker build -t canvasxpress-mcp-server:latest .

Initialize Vector Database

make init
# This creates ./vector_db/ directory with embedded examples

Start MCP Server

make run-http  # HTTP mode (daemon, background)
# Check status with: make logs

Configure Claude Desktop

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "canvasxpress": {
      "command": "docker",
      "args": [
        "exec",
        "-i",
        "canvasxpress-mcp-server",
        "python",
        "/app/src/mcp_server.py"
      ]
    }
  }
}

Restart Claude Desktop.

Test in Claude

Open Claude Desktop and try:

"Use the CanvasXpress tool to create a bar chart showing sales by region with blue bars and a legend on the right"

📖 Usage

Transport Modes

The MCP server supports two transport modes:

1. HTTP Mode (Default) - Network Access 🌐

make run-http   # Daemon mode (background)
make run-httpi  # Interactive mode (foreground)

• URL: http://localhost:8000/mcp

• Access: From anywhere on network/internet

• Clients: Multiple simultaneous connections

• Use cases:

  • Remote AI assistants (ChatGPT, Claude web)

  • Cloud deployment (AWS, GCP, Azure)

  • Team collaboration

  • Production services

2. STDIO Mode - Local Only 💻

make run

• Access: Same machine only

• Clients: Single local client (Claude Desktop)

• Use cases:

  • Claude Desktop integration

  • Local development

  • Private/offline usage

To switch modes, edit .env:

# For HTTP mode (default)
MCP_TRANSPORT=http

# For STDIO mode
MCP_TRANSPORT=stdio

CLI Client (mcp_cli.py)

Command-line interface for querying the HTTP MCP server with custom requests.

Basic Usage

# Simple query
python3 mcp_cli.py -q "Generate a bar graph with title 'hello, world'"

# With column headers
python3 mcp_cli.py -q "Scatter plot of expression over time" --headers "Time,Expression,Gene"

# Adjust LLM temperature
python3 mcp_cli.py -q "Create a clustered heatmap" --temperature 0.2

# Connect to remote server
python3 mcp_cli.py -q "Bar chart" --url http://myserver:8000

# JSON output (for piping)
python3 mcp_cli.py -q "Line chart" --json

# Config-only output
python3 mcp_cli.py -q "Line chart" --config-only

Inside Docker

# Run directly in container
docker exec -it $(docker ps -q) python3 /app/mcp_cli.py -q "Create a scatter plot"

# Or exec into container first
docker exec -it $(docker ps -q) /bin/bash
python3 mcp_cli.py -q "Pie chart showing market share"

Available Options

Note: The CLI client connects to the running HTTP MCP server. Make sure the server is running with make run-http first.

Available Tool

The server provides one tool: generate_canvasxpress_config

Parameters:

• description (required): Natural language description of visualization

• headers (optional): Column names from your dataset

• temperature (optional): LLM temperature, default 0.0

Example:

Description: "Create a scatter plot with log scale on y-axis, red points, and regression line"
Headers: "Time, Expression, Gene"
Temperature: 0.0

Supported Chart Types

Bar, Boxplot, Scatter, Line, Heatmap, Area, Dotplot, Pie, Venn, Network, Sankey, Genome, Stacked, Circular, Radar, Bubble, Candlestick, and 40+ more.

🔧 Configuration

Environment Variables

Configure your .env file with your BMS credentials:

# Azure OpenAI Configuration
AZURE_OPENAI_KEY=your_key_from_genai.web.bms.com
AZURE_OPENAI_API_VERSION=2024-02-01
LLM_MODEL=gpt-4o-global
LLM_ENVIRONMENT=nonprod

# MCP Server Configuration
MCP_TRANSPORT=http              # http (network, default) or stdio (local)
MCP_HOST=0.0.0.0               # HTTP: bind to all interfaces
MCP_PORT=8000                  # HTTP: port to listen on

Available Azure Models

BMS Proxy Details

• Endpoints URL: https://bms-openai-proxy-eus-prod.azu.bms.com/openai-urls.json

• Retry Logic: Automatically rotates endpoints on failures (429, connection errors)

• Max Retries: 3 attempts per request

• API Version: 2024-02-01

🛠️ Development

Makefile Commands

make help       # Show all commands
make build      # Build Docker image
make init       # Initialize vector database
make test-db    # Test vector database (inspect examples, search)
make run        # Start server (STDIO mode)
make run-http   # Start server (HTTP mode, daemon/background)
make run-httpi  # Start server (HTTP mode, interactive)
make stop       # Stop server
make logs       # View logs
make shell      # Open shell in container
make clean      # Remove container and image

Testing Utilities

Vector Database Testing

Test and inspect the Milvus vector database:

make test-db

What it tests:

• ✓ Database connection and collection info

• ✓ Row count (should be 132 examples)

• ✓ Sample data display (first 3 examples)

• ✓ Chart type distribution (30 unique types)

• ✓ Vector dimensions (1024 for BGE-M3)

• ✓ Semantic search with 3 sample queries

Example output:

✓ Collections: ['few_shot_examples']
✓ Row count: 132
✓ Total unique chart types: 30
✓ Vector dimension: 1024 (BGE-M3) or 1536 (OpenAI) or 768 (Gemini)

🔍 SEMANTIC SEARCH TEST
Query: 'bar chart with blue bars'
--- Result 1 (Similarity: 0.5845) ---
ID: 11
Type: Bar
Description: Create a bar graph with vertical orientation...

This is useful for:

• Verifying database initialization

• Understanding what examples are available

• Testing semantic search quality

• Debugging RAG retrieval issues

Python Usage Examples

Test the generator as a Python library:

# Set environment variables
source <(sed 's/^/export /' .env)

# Run examples
python3 examples_usage.py

See examples_usage.py and PYTHON_USAGE.md for detailed usage patterns.

Directory Structure

canvasxpress-mcp-server/
├── data/                       # Few-shot examples and schema
│   ├── few_shot_examples.json  # 132 examples (66 configs × 2 descriptions)
│   ├── schema.md               # CanvasXpress config schema
│   └── prompt_template.md      # LLM prompt template
├── src/                        # Source code
│   ├── canvasxpress_generator.py  # Core RAG pipeline + provider classes
│   └── mcp_server.py           # FastMCP server entry point
├── scripts/                    # Utility scripts
│   └── init_vector_db.py       # Local venv DB initialization
├── vector_db/                  # Vector database (auto-created)
│   └── canvasxpress_mcp.db     # Milvus database with embeddings
├── venv/                       # Python venv (if using local setup)
├── mcp_cli.py                  # CLI client for HTTP server
├── mcp_http_client.py          # HTTP client examples
├── test_vector_db.py           # Vector database testing utility
├── examples_usage.py           # Python usage examples
├── Dockerfile
├── Makefile
├── requirements.txt            # Full deps (with PyTorch/BGE-M3)
├── requirements-light.txt      # Lightweight deps (cloud only)
├── .env.example
└── README.md

Local Testing (without Docker)

For local development, use the virtual environment setup:

# Create venv (full or lightweight)
make venv        # Full (~8GB, includes BGE-M3)
make venv-light  # Lightweight (~500MB, cloud embeddings)

# Configure environment
cp .env.example .env
nano .env  # Add your API key, set EMBEDDING_PROVIDER if using venv-light

# Initialize database
make init-local

# Start server
make run-local   # HTTP mode
make run-locali  # STDIO mode

# Test with CLI
source venv/bin/activate
python mcp_cli.py -q "Create a bar chart"

See the "Local Development" section above for detailed instructions.

📊 Methodology

Based on peer-reviewed research (Smith & Neuhaus, 2024):

1. Embedding Model: BGE-M3 (1024d, local) or cloud alternatives (OpenAI 1536d, Gemini 768d)

2. Vector DB: Milvus-lite (local SQLite-based storage)

3. Few-Shot Examples: 132 examples (66 configs with human + GPT-4 descriptions)

4. Retrieval: Top 25 most similar examples per query

5. LLM: Azure OpenAI (BMS proxy) or Google Gemini

6. Retry Logic: Endpoint rotation on failures (BMS proxy pattern)

Performance (with BGE-M3 + Azure OpenAI):

• 93% exact match accuracy

• 98% similarity score

• Handles 30+ chart types

• Automatic failover across Azure regions

� Data Files

The data/ directory contains the files used for RAG and prompt generation:

Switching Few-Shot Files:

To use the expanded examples, create a symlink:

cd data/
mv few_shot_examples.json few_shot_examples_small.json  # Backup original
ln -s few_shot_examples_full.json few_shot_examples.json

Then reinitialize the vector database:

make init  # or make init-local for venv

Note: The expanded file is ~5MB vs ~40KB for the default. It provides better coverage but takes longer to embed on first run.

�🐛 Troubleshooting

Server won't start

# Check logs
make logs

# Verify .env file exists
ls -la .env

# Rebuild
make clean
make build
make init

Vector database errors

# Test the database first
make test-db

# If you see "Row count: 0" or errors, reinitialize:
sudo rm -rf vector_db/canvasxpress_mcp.db vector_db/milvus
make init

# Verify it worked
make test-db

Note: The vector database files are created by Docker with root ownership, so you need sudo to delete them.

Empty database after init

If make init shows success but make test-db reports 0 rows:

# The collection exists but is empty, force recreation
sudo rm -rf vector_db/canvasxpress_mcp.db vector_db/milvus
make init

This ensures the database is populated with all 132 examples.

Testing semantic search

# Run the test utility to see RAG in action
make test-db

# This will:
# - Show all 132 examples are loaded
# - Display chart type distribution
# - Run sample semantic searches
# - Verify embedding dimensions (1024)

API key issues

# Test API key and BMS proxy
docker run --rm --env-file .env canvasxpress-mcp-server:latest \
  python -c "import os, requests; print('Key:', os.environ.get('AZURE_OPENAI_KEY')); r = requests.get('https://bms-openai-proxy-eus-prod.azu.bms.com/openai-urls.json'); print('Proxy:', r.status_code)"

BMS Proxy Issues

# Check if you can reach BMS proxy
curl https://bms-openai-proxy-eus-prod.azu.bms.com/openai-urls.json

# If timeout, ensure you're on BMS network or VPN

Claude Desktop not connecting

1. Check config file path: ~/Library/Application Support/Claude/claude_desktop_config.json

2. Verify server is running: docker ps | grep canvasxpress

3. Restart Claude Desktop

4. Check Claude Desktop logs: ~/Library/Logs/Claude/

📚 Resources

• Original Research: Smith & Neuhaus (2024) - CanvasXpress NLP Preprint

• FastMCP 2.0: https://gofastmcp.com/

• CanvasXpress: https://www.canvasxpress.org/

• MCP Protocol: https://modelcontextprotocol.io/

• BGE-M3 Model: https://huggingface.co/BAAI/bge-m3

📝 License

Based on reference implementation (JOSS publication). MCP server implementation: MIT License

🤝 Contributing

Issues and pull requests welcome!

📧 Support

For issues related to:

• MCP Server: Open GitHub issue

• CanvasXpress: See https://www.canvasxpress.org/documentation.html

• MCP Protocol: See https://modelcontextprotocol.io/docs

Built with ❤️ using