Share Snowflake-managed MCP servers

As a provider, you can share a Snowflake-managed MCP server with other organizations on Snowflake, allowing consumers to access your data through third-party AI tools such as Claude, Gemini, and ChatGPT using the Model Context Protocol (MCP). When you share an MCP server with a data product, consumers can immediately use compatible AI tools to interact with your shared data without creating their own MCP server or configuring tools manually. You define exactly which tools and capabilities consumers can access, so all consumers get the same pre-configured tool interface for your data. For more information about creating MCP servers, see Create an MCP Server object.

As a consumer of a shared Snowflake-managed MCP server, you can connect from your preferred MCP-compatible AI tool without setting up your own MCP server. After you install a listing that contains a shared MCP server, see MCP server URL to connect and Set up OAuth authentication to configure access. For more information on consuming shared data products, see Use listings as a consumer.

Requirements

Sharing an MCP server requires the following:

  • An existing Snowflake-managed MCP server that you want to share.
  • All tools referenced by the MCP server must be in the same database as the MCP server.
  • All underlying objects referenced by the MCP server tools (such as Cortex Search Services, semantic views, or Cortex Agents) must also be shared.

Set an MCP server as shared

You can share your MCP servers as a provider through Provider Studio (for Snowflake Marketplace listings) or through Internal sharing (for Internal Marketplace listings).

To share an MCP server in Snowsight:

  1. Sign in to Snowsight.
  2. Do one of the following:
    • For a Snowflake Marketplace listing, in the navigation menu, select Marketplace » Provider Studio.
    • For an Internal Marketplace listing, in the navigation menu, select Data sharing » Internal sharing.
  3. Open the listing, and then add the MCP server to the share attached to the listing.

You can also set an MCP server as shared with a SQL statement. The following example adds the MCP server my_mcp_server to the share my_share:

GRANT USAGE ON MCP SERVER my_mcp_server TO SHARE my_share;

If your MCP server references underlying objects such as Cortex Search Services, semantic views, or Cortex Agents, you must also grant those objects to the share:

GRANT USAGE ON MCP SERVER my_mcp_server TO SHARE my_share;
GRANT SELECT, REFERENCES ON SEMANTIC VIEW my_sv TO SHARE my_share;
GRANT USAGE ON CORTEX SEARCH SERVICE my_css TO SHARE my_share;
GRANT USAGE ON AGENT my_agent TO SHARE my_share;

Identify shared MCP servers in Snowsight

After a consumer installs the listing, in the navigation menu, select Catalog » Database Explorer. The shared MCP server appears under the installed listing’s database, alongside the other shared objects. Providers can review the objects included in a share from the listing’s detail page.

Replication

Shared MCP servers support replication. Providers can enable listing auto-fulfillment (LAF) to replicate MCP servers to other regions, allowing consumers in different regions to access the shared MCP server.

Limitations

You can share MCP servers that contain the following tool types:

  • CORTEX_SEARCH_SERVICE_QUERY: Cortex Search Service tool
  • CORTEX_ANALYST_MESSAGE: Cortex Analyst tool
  • SYSTEM_EXECUTE_SQL: SQL execution
  • CORTEX_AGENT_RUN: Cortex Agent tool
  • GENERIC: tool for UDFs and stored procedures

When you share an MCP server, you must also share all the objects referenced by its tools, similar to sharing a semantic view with its referenced Cortex Search Services and tables.

The following limitations also apply:

  • All tools must be in the same database as the MCP server.
  • MCP servers with tools that reference stored procedures can’t be shared.
  • A SQL table function can be shared, but a Python user-defined table function can’t.
  • If you update a shared Snowflake-managed MCP server to use new tools (such as semantic views, Cortex Search Services, or functions), you must also grant those new tools to the share. New tools aren’t automatically added.

Cost considerations

In addition to any costs paid to the provider of the shared data product, consumers are billed for the following:

  • Input and output tokens used by AI model invocations through the shared MCP server.
  • Consumer’s warehouse usage for SQL query and tool execution.

For more information on costs paid to providers, see Pay for Snowflake Marketplace listings. For more information on Snowflake costs, see the Snowflake Service Consumption Table.