AI Agent from Production to Deployment

A 3-part project built with Google Cloud, showing how to go from a simple MCP server all the way to a GPU-accelerated AI agent, which is all deployed on Google Cloud Run.

Each part is self-contained with its own README, but they build on each other.

What is Built

Architecture Overview

Prerequisites

Before starting, make sure you have:

• A Google Cloud account with billing enabled (free trial available)
• The gcloud CLI installed and authenticated
• Python 3.13+
• uv — the Python package manager used throughout (install)

All parts can run in Google Cloud Shell, which comes with all of the above pre-installed. Click here to open it.

Getting Started

Clone the repo and navigate into it:

git clone https://github.com/YOUR_USERNAME/zoo-mcp-on-cloudrun.git
cd zoo-mcp-on-cloudrun

Then start with Part 1:

cd 1-mcp-server
cat README.md

Cost Estimate

Each part README includes a clean up section to delete resources and avoid ongoing charges.

Project at a Glance

Part 1 — Deploying a secure MCP server on Cloud Run

Building a Model Context Protocol server using FastMCP that exposes zoo animal data as tools. Deploying it to Cloud Run with authentication required, then connecting to it using Gemini CLI.

Concepts involved: MCP concepts, FastMCP, deploying from source on Cloud Run, IAM-based auth, service accounts.

→ Go to Part 1

Part 2 — Building and Deploying an ADK Agent that uses an MCP Server

Building a multi-agent zoo tour guide using Google's Agent Development Kit (ADK). The agent uses the MCP server from Part 1 as its toolset, augmented with the Wikipedia API for general knowledge. Deploying the agent to Cloud Run.

Concepts involved: ADK agents, SequentialAgent, MCPToolset, LangchainTool, state management, adk deploy.

→ Go to Part 2

Part 3 — Deploying an ADK agent to Cloud Run with GPU

Deploying a GPU-accelerated Gemma 3 model via Ollama on Cloud Run, then wiring it up to an ADK agent. Running elasticity tests with Locust to observe how both services handle load independently.

Concepts involved: GPU on Cloud Run, Ollama, LiteLlm, FastAPI + ADK, Locust load testing, autoscaling behavior.

→ Go to Part 3

Project Structure

ai-agent-prod-deployment/
├── README.md                        # You are here
├── .gitignore
├── 1-mcp-server/
│   ├── README.md
│   ├── server.py                    # FastMCP zoo server with 2 tools
│   ├── Dockerfile
│   └── pyproject.toml
├── 2-adk-agent/
│   ├── README.md
│   ├── zoo_guide_agent/
│   │   ├── __init__.py
│   │   └── agent.py                 # Multi-agent workflow (greeter → researcher → formatter)
│   ├── requirements.txt
│   └── .env.example
└── 3-gpu-agent/
    ├── README.md
    ├── ollama-backend/
    │   └── Dockerfile               # Gemma 3 270M via Ollama
    └── adk-agent/
        ├── production_agent/
        │   ├── __init__.py
        │   └── agent.py             # Gemma-powered conversational agent
        ├── server.py                # FastAPI server with ADK integration
        ├── Dockerfile
        ├── elasticity_test.py       # Locust load test
        ├── pyproject.toml
        └── .env.example