Learn how to access Snowflake-hosted AI models and data from external applications using two different approaches: the OpenAI SDK for simple chat interfaces and LangGraph with Model Context Protocol (MCP) for custom agent workflows.

You'll build two complete applications that demonstrate external access patterns for Snowflake:

App 1: Chat with GPT-5 - A Streamlit chat interface that uses the OpenAI SDK to access Snowflake-hosted models like GPT-5 and Claude.

App 2: LangGraph MCP Agent - A LangGraph agent that connects to Snowflake's MCP server, enabling AI-powered sales intelligence through conversation search, metrics analysis, and data queries.

Download the get-started-with-openai-sdk-and-managed-mcp-for-cortex-agents.zip file which contains:

Extract the zip file to your preferred location.

1. Open a SQL worksheet in Snowsight
   • Navigate to Projects → Worksheets in the left navigation
   • Click + Worksheet to create a new SQL worksheet
2. Open the setup.sql file from the extracted zip folder
3. Copy and paste the entire script into the worksheet
4. Run all statements by clicking Run All at the top right

What the setup script does:

Before setting up the MCP server, create the Cortex Agent through Snowsight's UI:

1. Navigate to AI & ML → Agents in the left navigation menu
2. Click + Agent in the top right corner
3. Configure the agent with the following settings:

Basic Information:

Add Tools (in this order):

Tool 1 - Cortex Search:

Tool 2 - Cortex Analyst:

4. Click Create Agent

Before proceeding, verify the agent works:

1. In the Agent interface, try asking: "What are the top performing sales reps?"
2. The agent should use Cortex Analyst to query the metrics
3. Try: "Find conversations about budget concerns"
4. The agent should use Cortex Search to find relevant conversations

Now that your Cortex Agent is created, set up the MCP server to expose it externally:

1. Open a new SQL worksheet in Snowsight
   • Navigate to Projects → Worksheets
   • Click + Worksheet
2. Open the mcp_setup.sql file from the extracted zip folder
3. Copy and paste the entire script into the worksheet
4. Run all statements by clicking Run All

What the MCP setup script does:

After running the script, you'll see output from SELECT SYSTEM$SHOW_OAUTH_CLIENT_SECRETS('SALES_MCP_OAUTH');:

You'll need a Programmatic Access Token (PAT) for both applications.

1. In Snowsight, navigate to Profile → Settings → Authentication
2. Click Generate new token
3. Select Single Role → SALES_INTELLIGENCE_ROLE
4. Copy the token (you won't be able to retrieve it later!)

Navigate to the extracted project directory:

Both applications share the same Python environment and configuration:

Create your environment file:

Edit .env and add all your credentials:

A simple chat interface that uses the OpenAI SDK to access Snowflake-hosted models.

Architecture:

Start the Streamlit application:

The app will open at http://localhost:8501

Try these sample queries:

Select from the dropdown menu:

A LangGraph agent that connects to Snowflake's Cortex Agent via Model Context Protocol (MCP) with OAuth 2.0 authentication.

Architecture:

The agent.py application demonstrates how to build a custom LangGraph agent that connects to Snowflake's MCP server. Key features:

1. OAuth 2.0 Authentication Flow:

2. MCP Protocol Integration:

3. LangGraph Workflow:

4. Interactive Chat Interface:

Start the interactive agent:

Your agent will:

1. Open browser for OAuth 2.0 authorization
2. Connect to Snowflake's MCP server
3. Discover available Cortex Agent tools
4. Initialize LangGraph workflow with Snowflake Cortex LLM
5. Start interactive session

Try these sample queries:

Search conversations:

Query metrics:

Combined analysis:

Direct SQL:

Congratulations! You've successfully built two applications that access Snowflake externally using different approaches. The openai_sdk.py application demonstrated using the OpenAI SDK for simple LLM inference, while the agent.py application showed how to connect your LangGraph agent to Snowflake's Cortex Agent via Model Context Protocol with OAuth 2.0 authentication.

These patterns enable you to:

Documentation: