Routes Snowflake operations from Claude to the Cortex Code CLI while keeping everything else local. The routing is LLM-based semantic analysis, not keywords, so it catches "show me my warehouses" without you saying "Snowflake." Ships with three approval modes (prompt/auto/envelope_only) and security envelopes (RO/RW/RESEARCH) that block destructive operations by default. Caches capabilities at session start so repeat queries skip the discovery overhead. Worth noting the approval prompts show predicted tools with confidence scores, which is more transparent than most agent handoff patterns. Requires Cortex CLI installed and a Snowflake connection configured.
npx skills add https://github.com/snowflake-labs/subagent-cortex-code --skill cortex-code