I built a personal "Agentic OS" that runs my DBA work — with approval gates, audit trails, and a $0.01 morning brief

Short summary

An Oracle DBA built a personal agent operating system with six skills, tiered tool permissions (read/write/spend/prod-touch), hashed single-use approvals, and per-run cost accounting down to four decimals. The trust model—not the AI—is the core product: every gated call halts with a dry-run preview, unattended cron runs queue for approval, and a localhost FastAPI dashboard tails the audit log live. A real test caught a prefix-rule bug that would have allowed an unprompted DELETE.

• •Six agent skills with manifest-driven tools, risk tiers, and cron scheduling
• •Approval gates use hashed, single-use argument matching—no replay attacks
• •Per-run SQLite/JSONL audit logs track every tool call, approval, and token cost

Generated with AI, which can make mistakes.