Enterprise MCP Agent Architecture

This repository contains a modular, Serverless AI architecture built completely independently of the Google Agent Development Kit (ADK). It leverages the pure google.genai SDK and the open-source Model Context Protocol (MCP), deeply integrated into the Gemini Enterprise (Vertex AI Agent Builder) UI.

🏗️ Architecture

1. MCP Tool Server (mcp_server.py): A standard FastMCP server that exposes tools (Wikipedia, Math, GitHub). Completely decoupled from any LLM logic.

2. Custom Gemini App (mcp_client_app.py): A standalone FastAPI Python backend client. It connects to the MCP server, processes user prompts using the Gemini 2.5 Flash model via the standard SDK, handles complex multi-tool execution loops, and returns clean JSON.

3. Client UI (client_ui.html): A frontend user interface for interacting with the Custom Gemini App.

4. Gemini Enterprise Playbook: The Google Cloud enterprise front-end can natively connect to the MCP server to directly expose these tools to a Vertex AI agent.

Related MCP server: Gemini Agent MCP Server

🚀 Deployment Instructions

1. Deploy the MCP Server

gcloud run deploy gemini-enterprise-mcp-tools \
  --source . \
  --region us-central1 \
  --allow-unauthenticated \
  --command "uvicorn,mcp_server:app,--host,0.0.0.0,--port,8080"

2. Deploy the FastAPI App Client

gcloud run deploy gemini-enterprise-mcp-app \
  --source . \
  --region us-central1 \
  --allow-unauthenticated \
  --command "uvicorn,mcp_client_app:app,--host,0.0.0.0,--port,8080"

(Note: Ensure you are authenticated and have the correct GCP project set up.)

🔗 How to Integrate with Gemini Enterprise (Vertex AI Agent Builder)

Follow these exact steps to connect the MCP Server directly to Vertex AI Agent Builder.

Step 1: Create the Agent

1. In the Google Cloud Console, search for Agent Builder (or Gemini Enterprise Agent Platform).

2. Go to Apps -> Create App.

3. Select Agent -> Build your own.

4. Give it a Display Name. CRITICAL: Set the Location to us-central1. Click Create.

Step 2: Configure the MCP Tool

1. In your new Agent dashboard, click Tools on the left menu.

2. Click Create Tool -> Model Context Protocol (MCP) (if available in UI, otherwise follow standard MCP integration steps provided by Google Cloud).

3. Connect the tool directly to the Cloud Run URL of the deployed gemini-enterprise-mcp-tools service.

4. Ensure your agent is configured to use the MCP tools for its generated responses.

Step 3: Configure the Playbook

1. Go to Playbooks on the left menu and click on the Default Generative Playbook.

2. Goal: Paste the following:

   You are an enterprise AI assistant. Your goal is to answer user questions using your MCP tools.

3. Available Tools: At the bottom of the page, ensure the MCP tools are enabled.

4. Click Save.

Step 4: Test & Publish!

1. Use the Preview chat on the right side of the screen to send a test message.

2. Once working, go to the Integrations tab to generate a Web Widget (Dialogflow Messenger) snippet, or connect it to Google Chat, Slack, or Microsoft Teams.

Cost: The entire backend is Serverless (Cloud Run scales to zero) and the Playbook has no hourly uptime fee. This architecture costs $0.00 when idle!