Azure AI Foundry Agent

This guide covers integrating the Zscaler MCP Server with Azure AI Foundry to create autonomous security agents powered by GPT-4o (or GPT-4) that can call 300+ Zscaler tools.

Video walkthrough

▶ Walkthrough: Zscaler MCP Server and Azure AI Foundry Agent (Wistia)

There are two ways to configure the Foundry agent:

Method	Description	Best For
API (CLI)	Create and manage the agent via azure_mcp_operations.py commands	Automation, CI/CD, scripted deployments
UI (Portal)	Configure the agent through the Azure AI Foundry portal	Visual setup, exploring capabilities

Both methods result in the same Foundry agent — they differ only in how you create and manage it.

Architecture

Data flow:

1. You send a natural language request (e.g., “List my ZPA application segments”)

2. Azure AI Foundry’s GPT-4o interprets the request and selects the right MCP tool

3. Foundry calls your deployed MCP server via the MCPTool integration

4. The MCP server authenticates to the Zscaler API and returns results

5. GPT-4o formats the response and presents it to you

Prerequisites

Before setting up the Foundry agent, you need:

1. A deployed Zscaler MCP Server — Deploy via Container Apps or VM first:

   cd integrations/azure
   python azure_mcp_operations.py deploy
   

   This gives you a public MCP URL (e.g., https://zscaler-mcp-xxx.azurecontainerapps.io/mcp).

2. An Azure AI Foundry project — See Creating a Foundry Project below.

3. Azure CLI authenticated — az login

4. Python packages — installed automatically when needed, or manually:

   pip install azure-ai-projects azure-identity
   

5. Environment variables (in your integrations/azure/.env file):

   AZURE_AI_PROJECT_ENDPOINT=https://<resource>.services.ai.azure.com/api/projects/<project>
   AZURE_OPENAI_MODEL=gpt-4o
   

Creating a Foundry Project

If you don’t already have an Azure AI Foundry project, follow these steps:

1. Open Azure AI Foundry

Go to ai.azure.com. You’ll see the Foundry home page with your recent projects (if any).

2. Create a New Project

Click Create new project from the left sidebar dropdown. Fill in:

• Project name: e.g., zscaler-mcp-server

• Foundry resource: Select an existing resource or create a new one (e.g., zscaler-mcp-server-resource)

• Region: Choose a region (e.g., East US 2)

• Subscription: Your Azure subscription

• Resource group: Select or create one

Click Create.

3. Wait for Project Creation

Foundry will take a few seconds to provision the project and its resources.

4. Project Ready

Once complete, you’ll see a confirmation dialog — click Let’s go.

5. Copy the Project Endpoint

The project overview page displays three key values:

• API key — Used for direct API access (not needed for the CLI method)

• Project endpoint — This is the AZURE_AI_PROJECT_ENDPOINT value you need

• Azure OpenAI endpoint — Used for direct OpenAI API access

Copy the Project endpoint URL. It looks like:

https://<resource>.services.ai.azure.com/api/projects/<project>

6. Add to Your .env File

Paste the endpoint into your integrations/azure/.env file:

AZURE_AI_PROJECT_ENDPOINT=https://zscaler-mcp-server-resource.services.ai.azure.com/api/projects/zscaler-mcp-server
AZURE_OPENAI_MODEL=gpt-4o

When running agent_create, if this variable is set in the .env file, it is automatically detected and you won’t be prompted for it.

Method 1: API (CLI) Integration

The CLI method uses azure_mcp_operations.py subcommands to manage the agent lifecycle.

Step 1: Deploy the MCP Server

If you haven’t already:

cd integrations/azure
python azure_mcp_operations.py deploy

Follow the interactive prompts to select deployment target, auth mode, and credentials. Note the MCP URL printed at the end.

Step 2: Create the Foundry Agent

python azure_mcp_operations.py agent_create

This command:

1. Reads the deployed MCP server URL and auth mode from the deployment state

2. Prompts for your Foundry project endpoint and model name (or reads from .env)

3. Builds authentication headers based on your MCP auth mode

4. Creates the agent in Azure AI Foundry with an MCPTool pointing to your MCP server

The script will prompt for Foundry configuration. You can load it from a .env file (option 1) or enter manually (option 2):

Azure AI Foundry Configuration
----------------------------------------

How would you like to provide Foundry configuration?
  1. Load from a .env file
  2. Enter manually

Select [1/2]: 1
Path to .env file: integrations/azure/.env

Tip

If AZURE_AI_PROJECT_ENDPOINT is set in your .env file, it is automatically detected. Otherwise, you’ll be prompted to paste the endpoint URL.

Once complete, the script displays a summary:

============================================================
  Foundry Agent Created
============================================================

  Agent Name:    zscaler-mcp-agent
  Version:       1
  Model:         gpt-4o
  MCP Server:    https://zscaler-mcp-xxx.azurecontainerapps.io/mcp

  Next steps:
    1. Start a chat session:
       python azure_mcp_operations.py agent_chat

Step 3: Chat with the Agent

Interactive session:

python azure_mcp_operations.py agent_chat

Single query:

python azure_mcp_operations.py agent_chat -m "List all ZPA application segments"

The chat session features interactive multi-turn conversation, MCP tool approval prompts (approve/deny each tool call), per-response token usage tracking, and an end-of-session summary (duration, messages, cumulative tokens).

In-chat commands:

Command	Description
help	Show available commands, usage tips, and example prompts
status	Show agent info, project endpoint, session duration, tokens, and messages sent
clear	Clear the terminal screen
reset	Reset the conversation context (clears response chain, token count, message count)
quit / exit / q	End the chat session and display a summary

Step 4: Manage the Agent

# Check agent status
python azure_mcp_operations.py agent_status

# Delete the agent
python azure_mcp_operations.py agent_destroy

# Delete without confirmation prompt
python azure_mcp_operations.py agent_destroy -y

Authentication Between Foundry and MCP Server

The Foundry agent authenticates to the MCP server using custom HTTP headers passed via MCPTool.headers. The headers vary by auth mode:

MCP Auth Mode	Headers Sent by Foundry
Zscaler	X-Zscaler-Client-ID + X-Zscaler-Client-Secret
API Key	X-MCP-API-Key
JWT / OIDCProxy	Not directly supported — use API Key or Zscaler mode for Foundry
None	No headers

Note

Foundry blocks the standard Authorization header for security. The MCP server’s custom header authentication (X-Zscaler-Client-ID, X-MCP-API-Key) works as a workaround.

Method 2: UI (Portal) Integration

You can also configure the Foundry agent through the Azure AI Foundry portal at ai.azure.com.

Step 1: Deploy the MCP Server

Same as Method 1 — you need a running MCP server with a public URL.

Step 2: Open Your Foundry Project

1. Go to ai.azure.com

2. Open your project

3. Navigate to Agents in the left sidebar

Step 3: Create an Agent

1. Click + New agent

2. Configure the agent:

   • Name: zscaler-mcp-agent

   • Model: Select your GPT-4o deployment

   • Instructions: Paste the agent instructions (see below)

3. Under Tools, add an MCP Tool:

   • Server URL: Your deployed MCP server URL

   • Headers: Add authentication headers based on your MCP auth mode

Step 4: Test in the Portal

Use the built-in chat interface in the Foundry portal to test the agent. Ask questions like:

• “What Zscaler services are available?”

• “List my ZPA application segments”

• “Show me ZIA firewall rules”

Agent Instructions

Use these instructions when creating the agent via the portal:

You are a Zscaler security assistant powered by the Zscaler MCP Server.

You have access to 300+ tools for managing the Zscaler Zero Trust Exchange:
- ZPA (Zscaler Private Access): Application segments, access policies, connectors
- ZIA (Zscaler Internet Access): Firewall rules, URL filtering, DLP, locations
- ZDX (Zscaler Digital Experience): Application health, device metrics, alerts
- ZCC (Zscaler Client Connector): Device enrollment, forwarding profiles
- ZTW (Zscaler Workload Segmentation): IP groups, network services
- EASM (External Attack Surface Management): Findings, lookalike domains
- ZIdentity: Users, groups, identity management
- Z-Insights: Web traffic analytics, cyber incidents, shadow IT

Always start by calling zscaler_get_available_services to discover which services
and tools are enabled on this server.

When asked to perform operations:
1. List/get the current state first
2. Confirm changes with the user before executing writes
3. For ZIA changes: remind the user to activate configuration after modifications

CLI Command Reference

Command	Description
python azure_mcp_operations.py deploy	Deploy MCP server (Container Apps or VM)
python azure_mcp_operations.py agent_create	Create Foundry agent pointing to deployed MCP server
python azure_mcp_operations.py agent_chat	Start interactive chat session
python azure_mcp_operations.py agent_chat -m "query"	Send a single query
python azure_mcp_operations.py agent_status	Show agent status
python azure_mcp_operations.py agent_destroy	Delete the agent
python azure_mcp_operations.py agent_destroy -y	Delete without confirmation
python azure_mcp_operations.py destroy	Tear down all Azure resources
python azure_mcp_operations.py status	Show deployment status
python azure_mcp_operations.py logs	Stream container/VM logs

Environment Variables

Variable	Required	Description
AZURE_AI_PROJECT_ENDPOINT	Yes	Foundry project endpoint URL
AZURE_OPENAI_MODEL	No	Model deployment name (default: gpt-4o)
ZSCALER_CLIENT_ID	Yes	Zscaler OneAPI client ID
ZSCALER_CLIENT_SECRET	Yes	Zscaler OneAPI client secret
ZSCALER_VANITY_DOMAIN	Yes	Zscaler vanity domain
ZSCALER_CUSTOMER_ID	Yes	Zscaler customer ID
ZSCALER_CLOUD	Yes	Zscaler cloud (e.g., beta, production)

Troubleshooting

“No deployment found” when running agent_create

The Foundry agent requires a deployed MCP server. Run deploy first:

python azure_mcp_operations.py deploy

Agent can’t reach the MCP server

Verify the MCP server is running and accessible:

python azure_mcp_operations.py status
curl -s https://<your-mcp-url>/mcp | head -1

“MCP approval requests do not have an approval” error

This happens when tool approval responses aren’t properly chained. The CLI handles this automatically via previous_response_id tracking. If using the portal, ensure you approve tool calls in sequence.

JWT/OIDCProxy auth modes with Foundry

Azure AI Foundry blocks the standard Authorization header in MCPTool.headers. If your MCP server uses JWT or OIDCProxy auth, consider:

1. Redeploying with api-key or zscaler auth mode

2. Or adding a reverse proxy that injects the auth token