Connect Snowflake to ChatGPT: Manage Data Assets and Cortex AI

You want to connect Snowflake to ChatGPT so your AI agents can manage virtual warehouses, orchestrate data pipelines, run Cortex AI inference, and audit access controls. If your team uses Claude, check out our guide on connecting Snowflake to Claude or explore our broader architectural overview on connecting Snowflake to AI Agents. Here is exactly how to do it using a Model Context Protocol (MCP) server.

Giving a Large Language Model (LLM) read and write access to a sprawling enterprise data platform like Snowflake is an engineering challenge. The mandate from executive leadership is clear: automate data operations and bring GenAI directly to the data layer. However, translating a conversational prompt into a secure, rate-limited, and context-aware Snowflake API request requires sophisticated infrastructure.

You can either spend weeks building, hosting, and maintaining a custom MCP server, or you can use a managed infrastructure layer that handles the boilerplate for you. This guide breaks down exactly how to use Truto to generate a secure, managed MCP server for Snowflake, connect it natively to ChatGPT, and execute complex data engineering workflows using natural language.

The Engineering Reality of the Snowflake API

A custom MCP server is a self-hosted integration layer that translates an LLM's tool calls into REST API requests. While Anthropic's open standard provides a predictable way for models to discover tools, the reality of implementing it against vendor APIs is painful. If you decide to build a custom MCP server for Snowflake, you are responsible for the entire API lifecycle.

Here are the specific integration challenges that break standard REST assumptions when working with the Snowflake API:

Compute Lifecycle Management

You cannot simply query data in Snowflake; you need active compute resources. Most endpoints require a running Virtual Warehouse. If a warehouse is suspended to save credits, your custom MCP server must instruct the LLM to explicitly call a resume endpoint, wait for the warehouse to provision, execute the primary task, and then ideally suspend the warehouse again. If your MCP server does not expose these lifecycle controls as distinct tools, your LLM will fail to execute queries or you will rack up massive compute bills.

Asynchronous Execution and Polling

Snowflake's REST API is deeply asynchronous for heavy data operations. Executing a statement often returns a query ID rather than the immediate result set. Your MCP server must implement polling logic to check the status of the execution, fetch the result partitions, and stitch them back together before returning the context to the LLM. If your custom server forces the LLM to handle the polling loop manually, the model will rapidly burn through its token context window and hallucinate a completion state.

Dynamic Context and RBAC

Snowflake's security model relies heavily on the active role context. An operation might fail not because the user lacks permission, but because the session is using the wrong role. Your MCP server needs to cleanly map tool definitions to explicit role and schema parameters, ensuring the AI agent explicitly states the context (e.g., database, schema, role) for every single call.

Strict Concurrency and Rate Limits

Snowflake enforces strict rate limits on API requests to prevent endpoint abuse. If your AI agent gets stuck in a loop and spams the API, Snowflake will return an HTTP 429 Too Many Requests error. It is critical to understand that Truto does not retry, throttle, or apply backoff on rate limit errors. When the upstream Snowflake API returns a 429, Truto passes that error directly to the caller.

Truto normalizes the upstream rate limit information into standardized headers (ratelimit-limit, ratelimit-remaining, ratelimit-reset) per the IETF specification. The LLM client or the calling agent is strictly responsible for implementing its own retry and exponential backoff logic. If your client fails to handle the rejection gracefully, the LLM assumes the tool call succeeded and will hallucinate a response.

The Managed MCP Approach

Instead of forcing your engineering team to build and host custom integration layers, Truto provides a managed MCP architecture. When a customer connects their Snowflake account, Truto automatically generates a set of MCP tools derived dynamically from the integration's documented API endpoints.

Every MCP server in Truto is scoped to a single integrated account. The server URL contains a cryptographic token that securely encodes which account to use and what tools to expose. The system acts strictly as a proxy layer - tool calls are executed against Snowflake in real time, and the results are returned directly to the LLM. Truto does not cache, retain, or train on your Snowflake data payloads.

How to Create the Snowflake MCP Server

You can generate an MCP server for Snowflake dynamically using either the Truto UI or the API.

Method 1: Via the Truto UI

This is the fastest method for internal IT admins and data engineers configuring their own AI workspaces.

1. Navigate to the Integrated Accounts page in your Truto dashboard and select your connected Snowflake instance.
2. Click the MCP Servers tab.
3. Click Create MCP Server.
4. Select your desired configuration (e.g., name the server, filter to specific tags, or restrict it to read methods only).
5. Copy the generated MCP server URL. You will use this URL to connect ChatGPT.

Method 2: Via the Truto API

If you are building a B2B SaaS product and need to programmatically provision MCP servers for your end-users, you can create them via the API. This endpoint validates that the Snowflake connection has tools available, generates a secure token, stores it in a highly available distributed key-value store, and returns a ready-to-use URL.

Endpoint: POST /integrated-account/:id/mcp

// Example Request
{
  "name": "Snowflake Data Ops MCP",
  "config": {
    "methods": ["read", "write", "custom"],
    "tags": ["warehouses", "tasks", "cortex"]
  },
  "expires_at": "2026-12-31T23:59:59Z"
}

Example Response:

{
  "id": "mcp-abc-123",
  "name": "Snowflake Data Ops MCP",
  "config": { 
    "methods": ["read", "write", "custom"],
    "tags": ["warehouses", "tasks", "cortex"]
  },
  "expires_at": "2026-12-31T23:59:59Z",
  "url": "https://api.truto.one/mcp/a1b2c3d4e5f67890"
}

Raw tokens are never stored in plain text. The random hex token is hashed before being stored, meaning the raw URL is only returned to you once during this creation step.

How to Connect the MCP Server to ChatGPT

Once you have the Truto MCP URL, connecting it to an LLM framework takes less than a minute.

Method A: Via the ChatGPT UI

For users operating directly inside the ChatGPT interface:

1. Open ChatGPT and navigate to Settings -> Apps -> Advanced settings.
2. Enable the Developer mode toggle (MCP support requires this feature flag).
3. Under the MCP servers / Custom connectors section, click to add a new server.
4. Enter a name for your connector (e.g., "Snowflake Production").
5. Paste the Truto MCP URL into the Server URL field.
6. Click Save.

ChatGPT will immediately ping the server, execute the protocol handshake, and load the available Snowflake tools.

(Note: If your team uses Claude Desktop, the process is similar. Go to Settings -> Integrations -> Add MCP Server, paste the URL, and click Add.)

Method B: Via Manual Config File

If you are running a local agentic framework, Cursor, or a headless LLM client, you can connect to the Truto MCP server using a standard JSON configuration file and the official Server-Sent Events (SSE) transport adapter.

{
  "mcpServers": {
    "snowflake_truto": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-sse",
        "https://api.truto.one/mcp/a1b2c3d4e5f67890"
      ]
    }
  }
}

6 Hero Tools for Snowflake Automation

Truto automatically generates dozens of tools for the Snowflake integration based on the underlying API resource schemas. Here are six high-leverage hero tools your AI agents can use to automate data ops.

1. Execute Cortex AI Inference

Tool: snowflake_cortex_inference_complete Snowflake Cortex is a fully managed service that provides access to industry-leading large language models directly inside your data warehouse. This tool allows your agent to pass unstructured text queries to Cortex models without moving data outside of Snowflake's security perimeter.

"Use Cortex to analyze the sentiment of the recent customer reviews stored in our database. Pass the prompt to the llama3-8b model via the cortex inference tool."

2. Resume Virtual Warehouses

Tool: snowflake_warehouses_resume Before executing heavy data transformations, an AI agent must ensure the necessary compute infrastructure is active. This tool resumes a suspended virtual warehouse.

"Check the status of the 'COMPUTE_WH' warehouse. If it is suspended, resume it so we can execute our data load tasks."

3. Suspend Virtual Warehouses

Tool: snowflake_warehouses_suspend Leaving warehouses running is the fastest way to burn through Snowflake credits. Agents can use this tool to explicitly suspend a warehouse, dropping all compute nodes and halting billing immediately after a job finishes.

"We have finished the nightly batch processing. Suspend the 'ETL_HEAVY_WH' warehouse immediately to prevent further billing."

4. Create Dynamic Tables

Tool: create_a_snowflake_dynamic_table Dynamic Tables are the modern way to build declarative data pipelines in Snowflake. This tool allows an agent to define a target table, specify a target lag (data freshness requirement), and provide the SQL query that populates it.

"Create a new dynamic table named 'daily_sales_rollup' in the 'analytics' schema. Use the 'COMPUTE_WH' warehouse, set the target lag to 1 hour, and use this SQL query to aggregate sales by region."

5. Execute Pipeline Tasks

Tool: snowflake_tasks_execute Tasks are used to schedule and run SQL statements or stored procedures. If a scheduled task fails or needs to be run out of band, an agent can manually trigger its execution.

"The marketing team needs their data refreshed early today. Execute the 'load_campaign_metrics' task in the 'marketing_db' schema."

6. Manage User Role Grants

Tool: create_a_snowflake_user_grant Managing access control is a major burden for IT admins. This tool allows an agent to grant a specific role to a Snowflake user programmatically.

"The new data scientist, Alice, needs access to the raw data schema. Grant the 'DATA_SCIENTIST_READ_ONLY' role to her user account."

To view the complete inventory of available Snowflake tools and their associated JSON schemas, visit the Snowflake integration page.

Workflows in Action

When you connect Snowflake to ChatGPT, you move beyond isolated API calls into multi-step, agentic workflows. Here is how different personas leverage these tools in production.

Scenario 1: The FinOps / Cost Optimization Agent

Data teams frequently struggle with runaway Snowflake compute costs. An AI agent can act as a FinOps assistant to manage warehouse lifecycles dynamically.

"I need to run an expensive historical backfill operation. Please resume the 'BATCH_PROCESSING_WH' warehouse, wait for it to be active, trigger the 'historical_sync' task, and then suspend the warehouse immediately when the task completes so we do not waste credits."

Step-by-step execution:

1. The agent calls snowflake_warehouses_resume with the target warehouse name.
2. The agent monitors the status, then calls snowflake_tasks_execute to run the specific job.
3. Upon confirming the job was accepted, the agent calls snowflake_warehouses_suspend to shut down the compute nodes.

Output: The user receives confirmation that the compute was spun up, the job was submitted, and the infrastructure was spun back down to save costs.

Scenario 2: The Data Governance Admin

Onboarding a new analyst requires provisioning a user account, assigning the correct roles, and ensuring they have access to the right compute resources.

"We just hired a new analyst named Bob. Create a new Snowflake user for him. Once created, grant him the 'ANALYST_ROLE' and ensure he has usage rights on the 'ANALYST_WH' compute pool."

Step-by-step execution:

1. The agent calls create_a_snowflake_user with the required parameters (name, password constraints, defaults).
2. The agent calls create_a_snowflake_user_grant to bind the new user to the ANALYST_ROLE.
3. The agent queries get_single_snowflake_user_by_id to verify the state of the account.

Output: The admin receives a complete summary of the provisioned user profile, the applied security grants, and a confirmation that the onboarding tasks are complete.

Security and Access Control

Giving an AI model access to your data warehouse requires strict security boundaries. Truto MCP servers provide several layers of access control out of the box:

• Method Filtering: You can restrict a server to specific operation types by configuring methods: ["read"]. This ensures the AI agent can only execute get or list operations, physically preventing it from dropping tables or deleting users.
• Tag Filtering: Integrations are organized by functional tags. You can restrict an MCP server to only expose tools tagged with "warehouses" or "tasks", completely hiding the user management tools from the AI's context.
• API Token Authentication: By enabling require_api_token_auth: true, the MCP server URL alone is no longer sufficient. The connecting client must also provide a valid Truto API token in the Authorization header. This prevents unauthorized users from using the tools even if the URL leaks in a log file.
• Automatic Expiration: You can set an expires_at timestamp. Once the datetime passes, the token is automatically purged from the distributed key-value store, immediately terminating the server's access. This is ideal for granting a contractor temporary access to troubleshoot a data pipeline.

Build Faster with Truto

Connecting ChatGPT to Snowflake shouldn't require your engineers to maintain pagination logic, manage OAuth lifecycles, or hand-code massive JSON schemas. By utilizing a managed MCP server, you turn your Snowflake data warehouse into a fully interactive API surface for LLMs in a matter of minutes. Your agents get the real-time read and write tools they need, and your engineering team avoids adding another custom microservice to their backlog.