VARA.ai's MCP setup provides a modular, tool-based AI workflow for automated customer refund decisions grounded in a policy knowledge graph. The system exposes three main MCP servers: db_verification_server for interacting with the order database, doc_server for processing and extracting data from invoices, and defect_analyzer for analyzing product defect imagery. Each server can be invoked as part of a coordinated processing pipeline via the MCP framework, enabling end-to-end reasoning with policy rules, database lookups, and evidence extraction. To use, deploy the Python MCP services, then connect client tooling or orchestrators to call the appropriate tools (e.g., list_orders_by_customer_email, process_invoice, analyze_defect_image) and compose results from the grounded reasoning chain. The Adjudicator Agent in the multi-agent system will produce a final refund decision with an explanation derived from the policy knowledge graph.

Prerequisites:

• Python 3.11+ and a virtual environment tool (venv/orconda)
• Access to required data services (PostgreSQL for orders, Neo4j for policy graph)
• Network access to any cloud resources used by the deployment (Cloud Run, etc.)

1. Clone the repository: git clone <repository-url> cd mcp_customer_support

2. Set up a Python virtual environment and install dependencies for the main services: python -m venv venv

   On Windows: venv\Scripts\activate

   On macOS/Linux: source venv/bin/activate

   pip install -r gmail-event-processor/requirements.txt pip install -r mcp_processor/requirements.txt

3. Configure environment variables (see notes below) and ensure database connections are reachable:

   • DB_VERIFICATION: PostgreSQL connection string for orders and customers
   • NE04J_URI: Neo4j Aura connection URI
   • NE04J_AUTH: Neo4j credentials
   • GOOGLE_APPLICATION_CREDENTIALS: Path to GCP service account key (if using Google services)
   • any other service credentials required by the deployment

4. Run the MCP servers (in separate terminals or as a managed service): python -m mcp_customer_support python -m mcp_customer_support.db_verification_server python -m mcp_customer_support.doc_server python -m mcp_customer_support.defect_analyzer

5. Verify the services are listening and reachable via your MCP orchestrator or CLI tools.

6. (Optional) If using Docker, build and run containers per submodule as needed, following the repository's dockerfiles.

Tips and common issues:

• Ensure PostgreSQL and Neo4j connections are reachable from the deployment environment; verify host, port, database name, and credentials.
• Keep environment variables secure; use Secret Manager or encrypted storage in production.
• If using Google Vision or Gemini services, ensure proper API keys and rate limits are configured.
• The MCP servers are designed to be invoked by an orchestration layer; to test locally, you can call the exposed tools from client code or via the MCP CLI once the servers are running.
• If you see connectivity errors, check network egress rules and firewall settings between services.
• For large policy graphs, monitor memory usage on the Neo4j Graph Engine and consider incremental loading or batching during graph construction.