SQL Guard is a natural-language-to-SQL agent whose generated SQL must clear a deterministic validator before it is allowed to run — a wrong answer becomes a refusal, not a confident wrong number. Aimed at teams that want AI querying over business data without trusting raw LLM output, it ships with typed boundaries, real error handling, guardrails, tracing, and a versioned eval harness. The full demo runs offline and reproducibly with zero API keys, falling back to a deterministic fake LLM and local SQLite.