Multi-Agent-Customer-Service-System-with-A2A-and-MCP

# Multi-Agent Customer Service System with MCP and A2A This project implements 5 requirements: - A fully functional MCP HTTP server exposing the five required database tools with SQLite storage and SSE streaming. - Multiple independent A2A agents (Router via LangGraph, Customer Data Agent, Support Agent, Billing Agent) each running their own FastAPI + JSON-RPC server. - A LangGraph-powered router agent that orchestrates multi-step workflows across specialists using real A2A requests. - Demo utilities and setup instructions for running the stack locally. ## Project layout ``` /mcp_server # FastAPI MCP server implementing tools/list and tools/call /agents/router # LangGraph router agent (A2A server) /agents/data # Customer Data Agent (calls MCP tools) /agents/support # Support Agent (non-billing support) /agents/billing # Billing Agent (optional specialist) /shared # Shared helpers (SQLite + A2A helpers) requirements.txt # Python dependencies ``` ## Prerequisites - Python 3.11+ - `pip install -r requirements.txt` ## Running the servers Open separate terminals for each service (all uvicorn processes bind to `0.0.0.0`): ```bash # 1) Start MCP server (tools/list, tools/call, SSE stream) uvicorn mcp_server.app:app --port 8000 --reload # 2) Customer Data Agent (calls MCP tools) uvicorn agents.data.main:app --port 8011 --reload # 3) Support Agent uvicorn agents.support.main:app --port 8012 --reload # 4) Billing Agent (optional) uvicorn agents.billing.main:app --port 8013 --reload # 5) LangGraph Router Agent uvicorn agents.router.main:app --port 8010 --reload ``` ## Demo script With all services running, execute: ```bash python demo.py "Need the history for customer 1" ``` The script will send a JSON-RPC `message/send` request to the router agent, which forwards to the data and support agents, then prints the aggregated response. ## MCP protocol support - `GET /tools/list` returns all tool definitions and JSON schemas. - `POST /tools/call` executes the requested tool. - `GET /events/stream` provides a Server-Sent Events stream of tool calls for observability. The MCP server uses SQLite and seeds example customers, tickets, and interactions on first start. ## A2A agent interoperability Each agent exposes: - Agent card at `/.well-known/agent.json` - JSON-RPC endpoint at `/rpc` (supports `message/send` + streaming) - Skills, capabilities, and metadata required by A2A clients (e.g., MCP Inspector, LangGraph flows). Router coordination leverages LangGraph for conditional routing, multi-step workflows (data→support), and escalation to billing when keywords like "payment" or "refund" appear. ## Test scenarios 1. Simple Query: "Get customer information for ID 5" Single agent, straightforward MCP call 2. Coordinated Query: "I'm customer 12345 and need help upgrading my account" Multiple agents coordinate: data fetch + support response 3. Complex Query: "Show me all active customers who have open tickets" Requires negotiation between data and support agents 4. Escalation: "I've been charged twice, please refund immediately!" Router must identify urgency and route appropriately 5. Multi-Intent: "Update my email to new@email.com and show my ticket history" Parallel task execution and coordination ## Notes - Environment variables `MCP_SERVER_URL`, `DATA_AGENT_RPC`, `SUPPORT_AGENT_RPC`, and `BILLING_AGENT_RPC` let you repoint agents to different hosts. - The shared SQLite database path defaults to `./database.sqlite` and will auto-initialize with seed data.