cortex-code

Cortex Code CLI Integration

Cortex Code is Snowflake's AI coding agent CLI (built on Claude Code). It has native access to Snowflake schema introspection, SQL execution, Streamlit deployment, Cortex Agent creation, dbt model generation, and Airflow DAG optimization.

When to Use Cortex Code vs snow CLI

Task	Tool
Run a SQL query	snow sql -q "SELECT ..."
List/manage Snowflake objects	snow object list ...
Deploy a Streamlit app	cortex
Create a Cortex Agent	cortex
Build a Snowflake Intelligence agent	cortex
Generate dbt models from schema	cortex
Explore schema, lineage, RBAC	cortex
Optimize Airflow DAGs	cortex
Complex multi-step Snowflake workflows	cortex

Rule of thumb: If it's a single SQL statement or object operation, use snow. If it requires Snowflake-native AI capabilities, schema awareness, or multi-step orchestration, use cortex.

Invocation Pattern

Non-interactive (single prompt, get result)

cortex -p "your prompt here" -c <connection> --dangerously-allow-all-tool-calls

Non-interactive with JSON output (for parsing)

cortex -p "your prompt here" -c <connection> --dangerously-allow-all-tool-calls --output-format stream-json

With a working directory (for file creation like Streamlit apps)

cortex -p "your prompt here" -c <connection> -w /workspace/group --dangerously-allow-all-tool-calls

Key Flags

Flag	Purpose
-p "prompt"	Non-interactive: pass prompt, print response, exit
-c <connection>	Select Snowflake connection from connections.toml
-w <path>	Working directory for file operations
--dangerously-allow-all-tool-calls	Skip permission prompts (required for non-interactive use)
--output-format stream-json	JSON output for scripting/parsing
-m <model>	Override AI model (default: auto)
--continue	Resume most recent conversation
-r <session-id>	Resume specific session

Connection Selection

Cortex Code reads connections from ~/.snowflake/connections.toml (same as snow CLI). Always specify -c <connection> to target the correct Snowflake account. Available connections depend on which are configured for your environment.

Example Workflows

Create a Cortex Agent

cortex -p "Create a Cortex Agent for inventory management that can answer questions about stock levels, reorder points, and supplier lead times using the INVENTORY schema in the ANALYTICS database. Deploy it to Snowflake Intelligence." -c apollo -w /workspace/group --dangerously-allow-all-tool-calls

Deploy a Streamlit App

cortex -p "Build an interactive Streamlit dashboard showing revenue by region with date filters using the SALES_MART.REVENUE table. Deploy it to Snowflake." -c apollo -w /workspace/group --dangerously-allow-all-tool-calls

Generate dbt Models

cortex -p "Create dbt staging models for all tables in the RAW.STRIPE schema. Follow the dbt style guide with proper materialization configs and tests." -c apollo -w /workspace/group --dangerously-allow-all-tool-calls

Schema Discovery

cortex -p "List all tables tagged with PII=TRUE and show their data lineage" -c apollo --dangerously-allow-all-tool-calls

Important Notes

• Cortex Code is a nested AI agent — it runs its own Claude session. This means additional LLM cost and latency on top of the outer agent. Use it surgically for workflows it's uniquely good at, not as a general-purpose Snowflake interface.
• Cortex Code creates files in its working directory. Always use -w to control where files land.
• Sessions are stored in ~/.snowflake/cortex/conversations/. Use --continue or -r to resume.
• For long-running tasks (Streamlit deploy, agent creation), expect 30-120 seconds of execution.