Skip to main content

Snowflake MCP Server

The Snowflake MCP server lets your AI tools run read-only SQL against your Snowflake account, so you can ask about your tables, columns, and data and get an answer from the warehouse instead of a guess.

It is backed by Snowflake's managed MCP server, which runs inside your own Snowflake account. Your data never leaves it.

Prerequisites​

  • A Snowflake environment whose first Snowflake connection is a development connection, that is, one whose name contains dev.
  • Key pair authentication on that connection. Snowflake accepts a key pair token and nothing else for MCP, so a password-authenticated connection cannot reach the server.
  • An MCP server object in your Snowflake account, created by an account administrator as shown below.
  • The Snowflake toggle turned on for your environment by an administrator, under Admin > Environments > AI Tools > MCP Servers.

Create the server in your Snowflake account​

The MCP server is an object in your own account. Datacoves does not create it and cannot see it until it exists. Pick the database and schema you want it in, and name the server DATACOVES.

create mcp server <database>.<schema>.DATACOVES from specification $$
{
  "version": 1,
  "tools": [
    {
      "title": "Read-only SQL",
      "name": "execute_sql",
      "type": "SYSTEM_EXECUTE_SQL",
      "description": "Run SELECT queries against Snowflake metadata and data.",
      "config": {
        "read_only": true,
        "query_timeout": 120,
        "warehouse": "<the warehouse your Datacoves connection uses>"
      }
    }
  ]
}
$$;

Then let the role your Datacoves connection uses reach it:

grant usage on mcp server <database>.<schema>.DATACOVES to role <your connection's role>;

read_only keeps the tool to SELECT queries, so the AI can read your data and metadata but cannot change anything.

Your account may hold other MCP servers. Datacoves picks the one named DATACOVES, and when the account holds exactly one MCP server it picks that one whatever it is called.

How it works​

One toggle, and nothing else to configure. Your workspace signs a token with the private key of your Snowflake connection, asks your account which MCP servers it holds, and serves that server to your AI tools from inside the workspace. Every request goes to Snowflake with a current token, so a conversation left open for hours keeps working.

The server is delivered to every AI tool in your workspace: Datacoves Copilot, GitHub Copilot in VS Code, the GitHub Copilot CLI, Claude Code (the extension and the CLI), OpenAI Codex and Snowflake Cortex. Adding it changes nothing else in your tools' MCP settings.

The server acts as you, through your connection's own Snowflake role, so it sees only what that role is allowed to see.

Until the server object exists, your AI tools don't list it. Once an administrator creates it, your workspace picks it up within five minutes, with no restart.

Use it​

Ask, for example:

Which columns does the customers table have, and how many rows are in it?

Count yesterday's rows in the raw orders table and compare them with the day before.

You can confirm it is connected in Datacoves Copilot under Settings > MCP Servers, where snowflake lists the execute_sql tool.

Learn more​