What can the Snowflake MCP server serve?
Per Snowflake's MCP documentation, it serves Cortex Analyst, Cortex Search and Cortex Agents as tools, plus custom tools and SQL.
| Tool type | What Snowflake says it is |
|---|---|
| CORTEX_AGENT_RUN | A Cortex Agent tool; Snowflake recommends this for governed business data questions |
| CORTEX_ANALYST_MESSAGE | A Cortex Analyst tool; exposed directly, it generates SQL and returns the statement to the client |
| CORTEX_SEARCH_SERVICE_QUERY | A Cortex Search Service tool |
| SYSTEM_EXECUTE_SQL | SQL execution, without Cortex Agent orchestration |
| GENERIC | A tool for user-defined functions and stored procedures |
| CODING_AGENT | A coding agent tool with no stored agent object |
Reviewed November 20, 2026. We have not connected a client to a live Snowflake account for this page.
How do you set it up?
The documentation describes creating the server as an object, then letting clients authenticate.
For business data, Snowflake recommends exposing one Cortex Agent as the client-facing tool, so the agent chooses between its Analyst, Search and custom tools. Compare with our Claude guides and AI implementation work.
- Go to the database and schema where the server should live.
- Run CREATE MCP SERVER with a specification (YAML) that lists each tool, its name, type, identifier and a precise description.
- Grant roles access to the server and, separately, to each tool.
- Connect an MCP client using Snowflake OAuth (the default), or External OAuth with your identity provider.
What are the security cautions?
Snowflake's page lists several, and they matter more than the setup.
- Data leaves Snowflake: tool results go to the MCP client and, through it, to its model provider under that party's terms. Decide what each tool may return.
- Per-tool permission: access to the server does not give access to its tools.
- OAuth over tokens: hard-coded tokens can leak; if you use a programmatic access token, give it the least-privileged role.
- Third-party servers: verify tools and descriptions before using servers from other accounts, to avoid tool poisoning or shadowing.
- Loops: an agent calling an MCP server that calls back into an agent can run up cost; Snowflake enforces a maximum recursion depth of 10.
- Hostnames: use hyphens, not underscores, in MCP server hostnames.
Frequently asked questions
What is the Snowflake MCP server?
A Snowflake-managed MCP server that lets AI agents retrieve data from your account, serving Cortex Analyst, Cortex Search, Cortex Agents, custom tools and SQL execution.
How does it authenticate?
Snowflake OAuth by default, with optional External OAuth against your identity provider. Snowflake recommends OAuth over hard-coded tokens.
Does an MCP client send my data to an AI provider?
Snowflake's page says tool results go to the MCP client and, through the client, to its model provider, under that party's terms. Decide what each tool may return.
Does connecting Claude or ChatGPT cost credits?
Queries run on warehouses and Cortex features use credits, as in the other guides. We have not measured a specific cost, and Snowflake documents a recursion limit to stop runaway loops.
Can you set this up for us?
Yes. We choose the tools, set roles and permissions and test with your AI client. Book a call to start.
Need help with Snowflake?
We help small and mid-size teams choose, size and run data tools. See our Snowflake guides and services, and the Power BI guides, and the Looker Studio guides, and the Claude guides, and the AI implementation.




