DataFrey brings text-to-SQL agent to Snowflake via MCP

DataFrey is a new MCP server targeting Snowflake, created by a data scientist who found Claude Code unusable for SQL tasks due to its lack of database context. Rather than following the native Snowflake MCP setup — which the author describes as painful enough to require "reconnecting with nature" — DataFrey provides a CLI wizard that handles DB connection setup and Claude Code plugin installation in one step. Users can then issue natural-language commands like `/db write dbt model to rank leads` directly inside Claude Code.

The author positions this as a practical alternative to Snowflake Cortex, which they say consistently fails to identify the correct tables.

For complex or ambiguous queries, DataFrey's `plan` tool activates a dedicated text-to-SQL agent that works in three stages: schema-aware RAG (including sample values) built during the initial DB connection, subagents that actively explore the data, and a planning layer. The author positions this as a practical alternative to Snowflake Cortex, which they say consistently fails to identify the correct tables. On security, the author acknowledges that database-as-MCP is a concern and says significant effort went into safer design; by default, the CLI requests SELECT permissions across all schemas, not just `information_schema`, based on the conviction that writing good SQL requires peeking at actual data.

The service is currently free and fully hosted by the author, subject to rate limits. The text-to-SQL planning component is intentionally closed-source, with plans to charge for planning calls above the free tier in the future. The author intends to open-source more of the project over time, noting it is currently deployed as a monolith.

Key facts

1. 01DataFrey is an MCP server for Snowflake only, built to give Claude Code proper database context for SQL tasks.
2. 02A CLI wizard handles DB connection setup and Claude Code plugin installation.
3. 03Users can issue natural-language commands such as `/db write dbt model to rank leads`.
4. 04A `plan` tool triggers a text-to-SQL agent that uses schema RAG, data-exploring subagents, and planning for complex queries.
5. 05The author says Snowflake Cortex is supposed to offer similar functionality but fails to find the right tables reliably.
6. 06By default, the CLI requests SELECT permissions on all schemas, not just `information_schema`, for better SQL accuracy.
7. 07The service is free and hosted by the author with rate limits; paid tiers for planning calls above the limit are planned.

Topics