Agentic Retail MCP System

A full-stack artificial intelligence application implementing a Model Context Protocol (MCP) server securely attached to a Firebase Firestore database, controlled via a Streamlit Chat interface and powered by OpenAI's gpt-4o.

System Architecture

This project is built using a decoupled Microservice pattern:

1. MCP Backend Server (src/server.py): A fastmcp Python server exposing Retail-specific tools ['check_inventory', 'calculate_elasticity', 'get_demand_forecast']. Connects directly to Google Cloud Firestore.
2. AI Streamlit Frontend (src/client.py): An asynchronous Web UI that communicates with your Backend Server using persistent Server-Sent Events (SSE). Parses tool inputs and loops them dynamically into OpenAI Function calls.

Prerequisites & Configuration

Before running any scripts, you must configure your .env file at the root of the project.

Create a .env file and define the following variables:

# Your OpenAI API Key for the GPT-4o Agent
OPENAI_API_KEY="sk-proj-YOUR-KEY-HERE"

# The absolute path to your Firebase Service Account JSON key
FIREBASE_CREDENTIALS_PATH="./firebase-key.json"

(If FIREBASE_CREDENTIALS_PATH is left blank, the system will attempt to utilize Google Cloud's Application Default Credentials natively).

Installation

This project utilizes uv as the lightning-fast python package dependency manager.

Run the following inside the root directory to ensure dependencies are mapped to Python >= 3.11:

uv init
uv add mcp pydantic firebase-admin python-dotenv openai streamlit

Database Initialization (Mock Data)

If you have a fresh Firebase Firestore instance, you can use the built-in seeding tools to populate your database with 100 sample retail products.

1. Generate the static CSV mock data structures:

   uv run python generate_data.py
   

2. Stream the generated CSV data directly into your Firebase NoSQL Database:

   uv run python seed_db.py
   

Running the Application

Because the Web Interface and the Tool Server communicate over an isolated local network loop, you must run both processes parallelly in two separate terminal tabs.

1. Boot up the Backend Tool Server

This stands up an ASGI Uvicorn app on port 8000 handling all tool execution requests natively via HTTP Server Sent Events.

uv run python src/server.py

2. Boot up the Streamlit Client

In a new terminal window, start the Streamlit conversational agent:

uv run streamlit run src/client.py 

A web browser tab will automatically open at http://localhost:8501. You can now ask questions like "What is the stock level of product P1002?" or "Analyze pricing elasticity projections for P1100".

Cloud Deployments (Google Cloud Run)

If you want to move your AI Agent and Backend infrastructure permanently to the cloud, I have configured advanced Dockerfiles designed explicitly for serverless deployment on Google Cloud Run.

Please refer to the complete DEPLOYMENT.md documentation guide for instructions on pushing both containers to the web.