QML-MCP: Quantum Machine Learning MCP Server

A Model Context Protocol (MCP) server for Quantum Machine Learning using Qiskit.

Features

• Quantum Circuit Execution: Run quantum circuits with configurable shots

• Quantum Kernel Computation: Compute quantum kernels for ML tasks

• Variational Quantum Classifier (VQC): Train quantum classifiers

• Model Evaluation: Evaluate trained quantum ML models

• Safety Limits: Configurable limits on qubits and shots

• Structured Logging: Comprehensive logging for debugging

• Error Handling: Detailed error messages with tracebacks

Installation

For development:

Requirements

• Python >= 3.10

• Qiskit >= 1.0.0, < 2.0.0 (Note: Qiskit Machine Learning 0.8.4 requires Qiskit 1.x)

• Qiskit Machine Learning >= 0.8.4

• MCP >= 0.9.0

Note on Qiskit Version: While Qiskit 2.0+ is available, Qiskit Machine Learning 0.8.4 (the latest stable version) requires Qiskit 1.x. This implementation uses Qiskit 1.4.5+ which provides all necessary quantum ML features.

Configuration

The server can be configured via environment variables:

• QML_MCP_QUANTUM_MAX_SHOTS: Maximum shots per circuit (default: 100000)

• QML_MCP_QUANTUM_MAX_QUBITS: Maximum qubits allowed (default: 10)

• QML_MCP_QUANTUM_DEFAULT_SHOTS: Default shots for circuits (default: 1024)

• QML_MCP_LOG_LEVEL: Logging level (default: INFO)

• QML_MCP_ENABLE_DETAILED_ERRORS: Include detailed error traces (default: true)

Usage

Running the Server

Available Tools

1. run_quantum_circuit

Execute a quantum circuit and get measurement results.

Parameters:

• qasm (required): Quantum circuit in QASM3 format

• shots (optional): Number of measurement shots (default: 1024)

Example:

2. compute_quantum_kernel

Compute quantum kernel matrix for ML tasks using ZZ feature map.

Parameters:

• train_data (required): Training data as 2D array

• test_data (optional): Test data as 2D array

• feature_dimension (optional): Number of features

Example:

3. train_vqc

Train a Variational Quantum Classifier.

Parameters:

• X_train (required): Training features as 2D array

• y_train (required): Training labels as 1D array

• feature_dimension (optional): Number of features

• max_iter (optional): Maximum optimization iterations (default: 100)

Example:

Returns a base64-encoded trained model.

4. evaluate_model

Evaluate a trained quantum ML model.

Parameters:

• model (required): Base64-encoded trained model

• X_test (required): Test features as 2D array

• y_test (optional): Test labels for accuracy computation

Example:

Testing

Run tests:

Run with coverage:

Project Structure

Safety and Limits

The server implements several safety mechanisms:

• Qubit Limits: Maximum number of qubits per circuit (default: 10)

• Shot Limits: Maximum measurement shots (default: 100000)

• Input Validation: All inputs are validated before processing

• Error Handling: Comprehensive error messages with optional tracebacks

License

MIT License - see LICENSE file for details.