Getting Started with Agent Service and Remote MCP Servers (Python)

This is a quickstart template to easily run an Azure AI Foundry Agent Service client and then add a custom remote MCP server to the cloud using Azure Functions Remote MCP. You can clone/restore/run on your local machine with debugging, and azd up to have it in the cloud in a couple minutes. The MCP server is secured by design using keys and HTTPS, and allows more options for OAuth using built-in auth and/or API Management as well as network isolation using VNET.

If you're looking for this sample in more languages check out the .NET/C# and Node.js/TypeScript versions.

Below is the architecture diagram for the Remote MCP Server using Azure Functions with Foundry Agent Service:

Prerequisites

• Python version 3.11 or higher
• Azure Functions Core Tools >= 4.0.7030
• Azure Developer CLI
• To use Visual Studio Code to run and debug locally:
  • Visual Studio Code
  • Azure Functions extension

Deploy Remote MCP Server to Azure

Run this azd command to provision the function app, with any required Azure resources including AI Foundry Agent Service, and deploy your code:

azd up

Note: You'll be prompted to specify an agentLocation during deployment. This must be one of the AI Foundry supported regions: westus, westus2, uaenorth, southindia, or switzerlandnorth. This location is used specifically for AI resources (AI Services, Search, Cosmos DB) and can be different from your main deployment location.

Additionally, API Management can be used for improved security and policies over your MCP Server, and App Service built-in authentication can be used to set up your favorite OAuth provider including Entra.

Connect to your remote MCP server function app from a client

Your client will need a key in order to invoke the new hosted SSE endpoint, which will be of the form https://<funcappname>.azurewebsites.net/runtime/webhooks/mcp/sse. The hosted function requires a system key by default which can be obtained from the portal or the CLI (az functionapp keys list --resource-group <resource_group> --name <function_app_name>). Obtain the system key named mcp_extension.

Foundry Agent Service Client

1. Change to the agent folder in a new terminal window:

   cd agent

2. Create a .env file based on the example provided. Copy the .env.example file:

   copy .env.example .env

3. Edit the .env file with your deployed function app details:

   # Azure AI Project Configuration
   PROJECT_ENDPOINT=https://your-agent-service-resource.services.ai.azure.com/api/projects/your-project-name
   MODEL_DEPLOYMENT_NAME=gpt-4.1-mini
   MCP_SERVER_LABEL=Azure_Functions_MCP_Server
   MCP_SERVER_URL=https://<your-funcappname>.azurewebsites.net/runtime/webhooks/mcp/sse
   USER_MESSAGE=Create a snippet called snippet1 that prints 'Hello, World!' in Python.
   
   # Required: Azure Functions extension key for MCP server authentication
   MCP_EXTENSION_KEY=your_mcp_extension_system_key_here

   Note: Replace the following values with outputs from your azd up deployment:

   • PROJECT_ENDPOINT: Your Azure AI Project endpoint (from azd deployment output)
   • <your-funcappname>: Your function app name (from azd deployment output)
   • your_mcp_extension_system_key_here: The mcp_extension system key obtained from the Azure portal or CLI

4. Install Python dependencies for the agent:

   pip install -r requirements.txt

5. Run the agent service:

   python main.py

   The agent will connect to your remote MCP server and execute the message specified in the USER_MESSAGE environment variable, demonstrating the integration between Azure AI Foundry and your deployed MCP server.

Connect to remote MCP server in MCP Inspector

For MCP Inspector, you can include the key in the URL:

https://<funcappname>.azurewebsites.net/runtime/webhooks/mcp/sse?code=<your-mcp-extension-system-key>

Redeploy your code

You can run the azd up command as many times as you need to both provision your Azure resources and deploy code updates to your function app.

Note

Deployed code files are always overwritten by the latest deployment package.

Clean up resources

When you're done working with your function app and related resources, you can use this command to delete the function app and its related resources from Azure and avoid incurring any further costs:

azd down

Helpful Azure Commands

Once your application is deployed, you can use these commands to manage and monitor your application:

# Get your function app name from the environment file
FUNCTION_APP_NAME=$(cat .azure/$(cat .azure/config.json | jq -r '.defaultEnvironment')/env.json | jq -r '.FUNCTION_APP_NAME')
echo $FUNCTION_APP_NAME

# Get resource group 
RESOURCE_GROUP=$(cat .azure/$(cat .azure/config.json | jq -r '.defaultEnvironment')/env.json | jq -r '.AZURE_RESOURCE_GROUP')
echo $RESOURCE_GROUP

# View function app logs
az webapp log tail --name $FUNCTION_APP_NAME --resource-group $RESOURCE_GROUP

# Redeploy the application without provisioning new resources
azd deploy

Debugging MCP server function locally

An Azure Storage Emulator is needed for this particular sample because we will save and get snippets from blob storage.

1. Start Azurite

   docker run -p 10000:10000 -p 10001:10001 -p 10002:10002 \
       mcr.microsoft.com/azure-storage/azurite

Note if you use Azurite coming from VS Code extension you need to run Azurite: Start now or you will see errors.

Run your MCP Server locally from the terminal

1. Change to the src/mcp_server folder in a new terminal window:

   cd src/mcp_server

2. Install Python dependencies:

   pip install -r requirements.txt

Note it is a best practice to create a Virtual Environment before doing the pip install to avoid dependency issues/collisions, or if you are running in CodeSpaces. See Python Environments in VS Code for more information.

1. Start the Functions host locally:

   func start

Note by default this will use the webhooks route: /runtime/webhooks/mcp/sse. Later we will use this in Azure to set the key on client/host calls: /runtime/webhooks/mcp/sse?code=<system_key>

Connect to the local MCP server from a client/host

Foundry Agent Service Client

The Foundry Agent Service is a cloud service that expects MCP tools that are also in the cloud (e.g. same VNET or on public internet). Proceed to the steps around deploying the Azure for Remote MCP.

MCP Inspector

1. In a new terminal window, install and run MCP Inspector

   npx @modelcontextprotocol/inspector

2. CTRL click to load the MCP Inspector web app from the URL displayed by the app (e.g. http://0.0.0.0:5173/#resources)

3. Set the transport type to SSE

4. Set the URL to your running Function app's SSE endpoint and Connect:

   http://0.0.0.0:7071/runtime/webhooks/mcp/sse

Note this step will not work in CodeSpaces. Please move on to Deploy to Remote MCP.

Source Code

The function code for the get_snippet and save_snippet endpoints are defined in the Python files in the src/mcp_server directory. The MCP function annotations expose these functions as MCP Server tools.

Here's the actual code from the function_app.py file:

@app.generic_trigger(arg_name="context", type="mcpToolTrigger", toolName="hello", 
                     description="Hello world.", 
                     toolProperties="[]")
def hello_mcp(context) -> None:
    """
    A simple function that returns a greeting message.

    Args:
        context: The trigger context (not used in this function).

    Returns:
        str: A greeting message.
    """
    return "Hello I am MCPTool!"


@app.generic_trigger(
    arg_name="context",
    type="mcpToolTrigger",
    toolName="getsnippet",
    description="Retrieve a snippet by name.",
    toolProperties=tool_properties_get_snippets_json
)
@app.generic_input_binding(
    arg_name="file",
    type="blob",
    connection="AzureWebJobsStorage",
    path=_BLOB_PATH
)
def get_snippet(file: func.InputStream, context) -> str:
    """
    Retrieves a snippet by name from Azure Blob Storage.
 
    Args:
        file (func.InputStream): The input binding to read the snippet from Azure Blob Storage.
        context: The trigger context containing the input arguments.
 
    Returns:
        str: The content of the snippet or an error message.
    """
    snippet_content = file.read().decode("utf-8")
    logging.info(f"Retrieved snippet: {snippet_content}")
    return snippet_content


@app.generic_trigger(
    arg_name="context",
    type="mcpToolTrigger",
    toolName="savesnippet",
    description="Save a snippet with a name.",
    toolProperties=tool_properties_save_snippets_json
)                   
@app.generic_output_binding(
    arg_name="file",
    type="blob",
    connection="AzureWebJobsStorage",
    path=_BLOB_PATH
)
def save_snippet(file: func.Out[str], context) -> str:
    content = json.loads(context)
    snippet_name_from_args = content["arguments"][_SNIPPET_NAME_PROPERTY_NAME]
    snippet_content_from_args = content["arguments"][_SNIPPET_PROPERTY_NAME]

    if not snippet_name_from_args:
        return "No snippet name provided"

    if not snippet_content_from_args:
        return "No snippet content provided"
 
    file.set(snippet_content_from_args)
    logging.info(f"Saved snippet: {snippet_content_from_args}")
    return f"Snippet '{snippet_content_from_args}' saved successfully"

Note that the host.json file also includes a reference to the experimental bundle, which is required for apps using this feature:

"extensionBundle": {
  "id": "Microsoft.Azure.Functions.ExtensionBundle.Experimental",
  "version": "[4.*, 5.0.0)"
}

Next Steps

• Add API Management to your MCP server (auth, gateway, policies, more!)
• Add built-in auth to your MCP server
• Enable VNET using VNET_ENABLED=true flag
• Learn more about related MCP efforts from Microsoft