fraud-detection

This MCP server provides a comprehensive Advanced Fraud Detection platform that combines multiple anomaly detection and graph-based techniques with agent-to-agent transaction protection. It exposes a suite of 24 tools across core fraud detection, agent-to-agent safeguards, and model management, enabling real-time transaction analysis, behavioral biometrics checks, network risk assessment, and explainable AI reasoning. You can call individual tools through the MCP client (for example, analyze_transaction, detect_behavioral_anomaly, assess_network_risk, explain_decision, classify_traffic_source, verify_agent_identity, and score_agent_reputation) to perform targeted analysis, verify agent credentials, or generate risk scores and explanations for fraud decisions. The system supports hybrid analysis by combining transaction data, behavioral patterns, and network relationships to produce a holistic risk assessment and actionable recommendations for intervention.

Prerequisites:

• Python 3.10 or newer
• Git
• Access to a Python virtual environment (recommended)

Step-by-step installation:

1. Clone the repository:

git clone https://github.com/marc-shade/fraud-detection-mcp
cd fraud-detection-mcp

2. (Optional) Create and activate a virtual environment:

python -m venv fraud_env
source fraud_env/bin/activate  # On Windows: fraud_env\Scripts\activate

3. Install dependencies:

pip install -r requirements.txt

4. Install the package (if required by the project layout):

python setup.py install

5. Start the MCP server using the provided configuration (adjust paths as needed):

# Example using the path to your Python environment and server script
python server.py

6. If you use the Claude Code Integration snippet, ensure the env vars point to your model/data locations and log level is set appropriately, then integrate into your Claude Desktop configuration as shown in the README snippet.

Tips and common configurations:

• Ensure FRAUD_DETECT_MODEL_PATH points to your trained models and feature stores. Update FRAUD_DETECT_LOG_LEVEL to control verbosity (INFO, DEBUG, WARNING).
• Use a Python virtual environment to isolate dependencies.
• If running in production, consider using a process manager (e.g., systemd, supervisord) to keep the MCP server running and to manage restarts.
• The MCP tools cover real-time analysis, batch processing, and model management; you can enable/disable tools via your client or configuration as needed.
• For deployment across environments, pin exact library versions in requirements.txt to maintain reproducibility.
• When debugging, start with health_check or get_model_status to verify the server and models are loading correctly before calling analysis tools.