MCP Snowflake Server
vManaged MCP server (GA 2025-11-04); snowflake-labs-mcp 1.x (deprecated)Snowflake (Official)
Official Snowflake MCP server, GA since 2025-11-04 as a fully managed in-account endpoint (CREATE MCP SERVER object). Exposes Cortex Analyst NL-to-SQL over semantic views, Cortex Search, Cortex Agents, SQL execution with configurable read-only mode, and custom UDF/procedure tools. OAuth 2.0 or Programmatic Access Token auth with per-tool RBAC; masking and data policies apply. The self-hosted snowflake-labs-mcp (PyPI) was deprecated in May 2026 in favor of the managed server.
Trust Vector Analysis
Dimension Breakdown
๐Performance & Reliability+
Endpoint availability and GA-status review
NL-to-SQL capability review; semantic-view grounding improves accuracy but accuracy still depends on semantic view quality
Retrieval capability assessment based on Cortex Search service documentation
SQL tool feature review
Failure-mode and limit documentation review
๐ก๏ธSecurity+
Authentication mechanism review
Access control model review; per-tool privilege separation is stronger than most MCP servers evaluated
Write-surface assessment; read-only mode plus RBAC scoping mitigates but writes to production data remain possible when enabled
Prompt-injection exposure analysis; governed perimeter constrains impact but the generic MCP injection vector via returned data is unmitigated
Credential handling review
๐Privacy & Compliance+
Data residency review
Platform certification review
Data flow analysis
Governance enforcement review
๐๏ธTrust & Transparency+
Documentation completeness review
Source availability review; scored down because the actively supported path is proprietary and the open-source path is end-of-life
Audit trail assessment
Community engagement analysis
โ๏ธOperational Excellence+
Setup complexity assessment
Scalability review
Cost analysis for agent-driven workloads
Ecosystem integration review
Operational burden assessment
- +Fully managed, GA MCP endpoint inside the Snowflake account โ no infrastructure to run
- +Per-tool RBAC: server access alone does not grant tool access; masking and data policies enforced server-side
- +OAuth 2.0 security integrations or revocable PATs โ no passwords or connection strings in client config
- +Cortex Analyst NL-to-SQL grounded in semantic views reduces hallucinated queries
- +SQL execution tool with configurable read-only mode, query timeout, and warehouse selection
- +Full audit trail via Snowflake query history and access history
- !Managed server is closed-source; the open-source snowflake-labs-mcp is deprecated (final release 2026-05-15)
- !Query and search results are exposed to the connected LLM provider
- !When the SQL tool's read-only mode is off, agents can write to production data within the role's privileges
- !OAuth sessions use only the user's DEFAULT_ROLE; secondary roles unsupported
- !Response truncation at 250 KB and 50-tool cap can constrain large workloads
- !Cortex Analyst supports semantic views only (not semantic models); not available in government regions
- !No documented result-wrapping or other prompt-injection countermeasure for untrusted data in results
Use Case Ratings
code generation
Good for generating SQL and Snowflake object DDL against real schemas
customer support
Cortex Search over support content works well; guard customer PII exposure
content creation
Limited direct fit; useful for data-backed content via Cortex Search
data analysis
Primary use case: governed NL-to-SQL over semantic views plus warehouse-scale SQL
research assistant
Strong retrieval over enterprise documents with Cortex Search and Knowledge Extensions
legal compliance
Server-side masking, RBAC, and audit history support compliance workflows
healthcare
HIPAA-certified platform helps, but LLM exposure of results needs careful policy design
financial analysis
Governed access to financial marts with masking policies; strong audit trail
education
Good for teaching governed analytics, though account setup is enterprise-oriented
creative writing
Not a fit beyond retrieving reference data for narratives