MCP Crash Course: Build a Conversational Multi-Agent System

Set up a basic MCP application with a client and server, connect via std io or http protocols, and prepare four-file VSCode project including resource file and a requirements file.

Create an MCP server with a tool that sums two integers, a resource that returns bill to pay file content, and a prompt to format results, then test client communication.

Create a MCP client application that connects to the server, lists available tools, resources, and prompts, and executes a tool. The client tests two protocols—ss and stdio—within one function.

Practice testing a conversational multi-agent system by running the MCP server and client via stdio or http with SSL. Learn to activate virtual environment and explore tools, resources, and prompts.

Learn to build a multi-agent MCP application with an LM model, featuring three agents—reception, sales, maintenance—and six tools, including handoff descriptions for agent handoffs.

Explore the car dealership relational database schema, including vehicles, sales, sellers, dealerships, customers, cities, and states, and learn to set up and query the database for model-ready data.

Create the nova drive motors database on your local Postgres by running the database SQL script to build states, cities, dealerships, customers, sellers, vehicles, and sales for the MCP application.

Run basic sql queries to inspect the main tables, revealing vehicles with name, type, price, and timestamps, plus dealerships, cities, sellers, and customers to support maintenance scheduling.

Set up the MCP multi-agent project by creating the client and server Python files, a requirements file, and an environment file for variables like the OpenAI API. Activate the virtual environment, install the dependencies (streamlit 5.0, python-dotenv, psycopg2, MCP 1.9.4, OpenAI library, and OpenAI agents 0.0.14), and prepare the application to run with these libraries.

Configure a dot env file with environment variables, including OpenAI API key and Postgres database details (host, port 5432, db name nova drive, user, password) for the MCP server.

Build a fast MCP server, load env vars, connect to Postgres, and implement reusable tool templates for querying vehicles, dealerships, sellers, and customer info, plus scheduling visits.

Build a client app with streamlit, using a session store to track messages, agents, and the active model, and configure three agents with different lm models.

Learn to build a Streamlit-based conversational ui with a centered image, a three-column layout, and session state storing user, assistant, and function call interactions.

Create MCP agents—reception, maintenance, and sales—configuring each with name, model, hand-off description, and instructions. Assign them to the Streamlit session state and set the reception agent as current to route.

Implement the chat resolution function that runs on each user message, orchestrating MCP servers and agents to generate a response while updating session history and the current agent.

Students test three customer scenarios—buying a vehicle, scheduling maintenance, and scheduling a test drive—in a conversational multi-agent app, using tools to fetch available vehicles and schedule visits.

Use the Streamlit app to schedule a car maintenance by retrieving customer data, selecting a vehicle, and setting service details, date, and time with the model's instruction-driven behavior.

Test case demonstrates interacting with a multi-agent system to find available vehicles, select a Chicago dealership, and successfully schedule a test drive.