isaacwasserman/mcp-snowflake-server

Purpose and Entry Point

This is a Python Model Context Protocol (MCP) server that gives MCP clients such as Claude Desktop controlled access to Snowflake SQL, metadata, and schema/insight resources. The executable server and MCP tool/resource registration are centered in src/mcp_snowflake_server/server.py; package entry-point metadata is expected in pyproject.toml.

Runtime and Deployment

The server runs as a local stdio MCP process, commonly launched through uv/uvx, and can also be containerized through Dockerfile. It targets Snowflake as its external system of record. Credentials and connection defaults can come from CLI/environment configuration or named TOML connections (example_connections.toml). runtime_config.json supplies runtime exclusion policy for database/schema/table discovery.

MCP Server Boundary

src/mcp_snowflake_server/server.py is the protocol-facing boundary. It parses operational flags such as connection selection, write enablement, logging, and tool exclusion; creates the service dependencies; then exposes MCP tools and URI resources. MCP clients never access the Snowflake connector directly.

Snowflake Access Layer

src/mcp_snowflake_server/db_client.py encapsulates Snowflake connection construction, query execution, and metadata discovery. Server handlers delegate read queries, permitted mutations/table creation, and database/schema/table/column inspection to this layer. This isolates Snowflake connector behavior, authentication modes, and result handling from MCP concerns.

Query Safety and Write Control

Read operations are available through the query tool path in src/mcp_snowflake_server/server.py. Write-capable tools are conditionally registered only when explicit write permission is enabled. src/mcp_snowflake_server/write_detector.py classifies or validates SQL intent, forming the key policy boundary that separates SELECT access from allowed INSERT, UPDATE, DELETE, and table-creation operations.

Schema Context Flow

Metadata requests flow MCP handler → database client → Snowflake → serialized tool response. When prefetching is enabled, discovered table schemas are additionally transformed into context://table/{table_name} resources. Filtering configured through runtime_config.json limits what catalog objects enter either tool results or precomputed context.

Insight Memo and Serialization

The server maintains an in-process, dynamically updated memo://insights resource; append_insight updates that memo and triggers resource-visible state changes in src/mcp_snowflake_server/server.py. src/mcp_snowflake_server/serialization.py provides the response boundary for converting Snowflake-native values and rows into MCP-safe structured results.