DataFrey – MCP Server for Snowflake with Text-to-SQL Agent

TL;DR

TL;DR: DataFrey is a Python MCP server that connects Snowflake to Claude Code, adding a text-to-SQL planning agent so you can query databases in plain English.

What Is DataFrey?

DataFrey is an open-source MCP server specifically for Snowflake, built by a data scientist who found Claude Code struggled with SQL without database context. It ships as a Python monorepo with three packages:

• datafrey-cli — interactive CLI wizard for auth, DB connection, and Claude Code plugin setup
• datafrey-mcp — the MCP server bridge
• datafrey-api — shared Pydantic models

The core value is the plan tool: when you ask a vague question, DataFrey runs a separate text-to-SQL agent that uses schema RAG, subagents for data exploration, and explicit planning to produce a reliable query.

Setup Workflow

Prerequisites

• Python 3.13 or higher
• A Snowflake account (user with appropriate permissions)
• Claude Code installed

Step 1: Install

Install via pip:

pip install datafrey

Or with uv:

uv tool install datafrey

Step 2: Authenticate and Connect

Run the interactive CLI:

datafrey

This starts a wizard that handles:

1. DataFrey login
2. Snowflake connection (interactive, no manual config)
3. Schema index sync
4. Claude Code plugin installation

Step 3: Connect to Claude Code

After installation, configure the Claude Code client:

datafrey client claude

Or for other MCP-compatible editors:

datafrey client cursor
datafrey client mcp   # generic MCP config block

Step 4: Query Your Database

In Claude Code, use natural language:

/db write dbt model to rank leads

For complex questions, the plan tool kicks in automatically to work through the query step by step.

Deeper Analysis

How the Planning Agent Works

The plan tool is the distinguishing feature. When you ask a complex question:

1. Schema RAG — DataFrey indexes your Snowflake schema (tables, columns, sample values) during connection setup
2. Subagent exploration — spawns lightweight agents to explore relevant tables before building the query
3. Planning step — breaks down the question into sub-queries, executes them, and synthesizes results

This differs from raw text-to-SQL in that it iterates rather than generating a single query and hoping it matches intent.

Security Design

The author notes that database-as-MCP is a sensitive pattern. By default, the CLI prompts for SELECT-only permissions. The architecture separates the MCP bridge from direct query execution, giving you control over what actions the agent can perform.

Package Architecture

The monorepo uses uv workspaces with four packages:

| Package | Purpose | | --- | --- | | datafrey-cli | CLI wizard, auth, DB connection | | datafrey-api | Shared Pydantic models | | datafrey-mcp | MCP server implementation | | datafrey-mock | Local dev mock API |

Only Snowflake is supported at launch, but the package structure suggests multi-database expansion is planned.

Practical Evaluation Checklist

• Python 3.13+ required (not 3.12 or 3.11)
• uv workspace monorepo — uv is the expected package manager
• Interactive CLI wizard means zero manual config for Snowflake connection
• Claude Code plugin installs as an MCP skill, not a native integration
• Schema index is built locally during datafrey index — no cloud processing of your schema
• No star history to evaluate (fresh repo, 0 stars)

Security Notes

• Default CLI behavior requests SELECT-only permissions — you can escalate manually
• Credentials stored via the CLI wizard, not hardcoded
• Schema index is local; your database structure is not sent to an external service
• For production use, review Snowflake role permissions before connecting

FAQ

Q: Does it support databases other than Snowflake? A: No, only Snowflake at launch. The repo structure is designed for expansion, but no other database drivers are included yet.

Q: How does the schema index work? A: During datafrey index, DataFrey reads your Snowflake schema metadata (tables, columns, types) and builds a local RAG index. This index is queried by the planning agent before generating SQL.

Q: Do I need a DataFrey cloud account? A: The CLI wizard requires a DataFrey login for authentication. The query execution itself runs against your own Snowflake instance.

Q: Can I use this with editors other than Claude Code? A: Yes. The datafrey client mcp command outputs a generic MCP config block you can use with any MCP-compatible editor. Cursor is explicitly supported.

Q: Is it open source? A: Yes, Apache-2.0 license. Source is on GitHub at github.com/datafrey-ai/datafrey.

Conclusion

DataFrey solves a real problem: Claude Code writing SQL blind, without schema context. The CLI wizard handles the annoying Snowflake connection setup, and the planning agent adds a layer of reliability for complex queries. Fresh repo with zero stars means it’s early — worth watching if Snowflake query context is a pain point for your team.

Source and Accuracy Notes

• Project page: datafrey.ai (unavailable during verification — DNS resolution failed)
• Documentation: docs.datafrey.ai (unavailable during verification)
• Source repository: github.com/datafrey-ai/datafrey
• License: Apache-2.0 (verified via GitHub API license.spdx_id)
• HN launch thread: news.ycombinator.com/item?id=47853007
• Source last checked: 2026-07-25 (commit 9a3f1b7, pushed 2026-07-11)