Snowflake Connection

A Kizen plugin that connects an Agentic Workflow to an external Snowflake database for real-time data lookups. It talks to Snowflake over the SQL REST API v2 and authenticates with a Programmatic Access Token (PAT).

The plugin ships two automation steps:

• Read Data (snowflake_read) runs read-only queries. A pre-flight check rejects anything that isn't a SELECT, so a workflow can't write by accident.
• Write Data (snowflake_write) runs INSERT, UPDATE, DELETE, and other DML/DDL. There is no guardrail, so validate any query built from user input before it runs.

Setup

Both steps read a secret named snowflake_connection. Its value must be valid JSON.

For a single environment, use a flat object:

{
  "warehouse": "COMPUTE_WH",
  "schema": "PUBLIC",
  "account": "myorg-account123",
  "pat": "ver:1-hint:abc123..."
}

For multiple environments, nest each one under its own key:

{
  "production_db": {
    "warehouse": "COMPUTE_WH",
    "schema": "PUBLIC",
    "account": "myorg-account123",
    "pat": "ver:1-hint:abc123..."
  },
  "development_db": {
    "warehouse": "COMPUTE_WH",
    "schema": "PUBLIC",
    "account": "myorg-account123",
    "pat": "ver:1-hint:abc123..."
  }
}

With a nested secret, set the Connection Secret Tag input to the key you want (for example production_db).

Inputs

Input	Type	Required	Description
Database	string	yes	Snowflake database to query
Query	string	yes	SQL statement to run
Return Single Value	boolean	yes	Expect a single cell and return it as a string. Fails if the result isn't exactly one row and one column.
Connection Secret Tag	string	no	Key to pick an environment from a nested connection secret

Outputs

Output	Type	Step	Description
Result	string	both	Query results. A single value, or the string form of the rows for a multi-row result.
Result Status	string	write only	Summary of what happened: the DML row counts when Snowflake returns stats, otherwise a row-count description.

Good to know

• Results come back as strings. Parse JSON or numeric types downstream if you need them.
• Queries block until they finish or hit the 30 second timeout. Long-running queries will time out.
• Only the first page of results is returned. Add a LIMIT for large result sets.
• Queries aren't parameterized. Build SQL from external input carefully to avoid injection.