Give an agent tools without making the model the authority.

This TypeScript + Postgres starter puts deterministic code and database rules around tool use:
• allow / deny / require-approval policy
• human approval with expiry and freshness checks
• hard call, cost, duration and retry budgets
• idempotent execution receipts so retries cannot duplicate side effects
• Supabase/Postgres RLS and current-membership checks
• append-only audit evidence and durable failure records
• real adversarial + concurrency tests
• operator console and safe reference agent

The bundled demo proves the boundary: reads can be allowed, destructive actions denied, writes held for human approval, approved work executed once, duplicate retries replayed safely, and over-budget work stopped before provider execution.

Includes the tested v1.0 source ZIP, migration + seed, tests, architecture/security/setup/support docs, and commercial source license. No model API key is required for the demo.

You connect your own auth context, real tools, provider SDKs, deployment, validation rules and reconciliation for external side effects. This is a tested governance starter, not a claim that arbitrary downstream tool code becomes automatically safe.