We let agents spend real money on servers: the guardrail stack that survived testing
Faye-Vozora · reddit · 2026-10-14
Small VPS host Vozora exposed their API via MCP so agents can order, resize, and destroy servers. The first design didn't survive contact with a real agent; here's what shipped:
- Price echo: the create call must contain the price the agent quoted the human; mismatch with live price means refusal (catches the most cases)
- Idempotency: a test agent retried a failed create three times — without a request ID that's three servers
- Risk-classed tools: spend, destroy, and reveal operations force human confirmation
- Per-key caps the agent can't raise: 5 orders/hr, £20/hr, £100/day
- Scoped keys: a monitoring agent physically can't order anything
What failed: trusting the model to "be careful" (meaningless under retries) and any limit stored where the agent can modify it. Key advice: if agents touch anything billable, design for the retry-loop failure mode first — it's boring and it's the one that actually happens.
More from coding & agent
- After 3 years, HTMS ships first client project compiling HTML into JavaScript — CtrlAltDwayne · 2026-10-14
- Codex tells its human to back off while it takes over a Premiere export — joonasvirtanen · 2026-10-14
- Dan Vega's dev2next talk: a 68-slide roadmap from Claude Code beginner to power user — therealdanvega · 2026-10-14
- Dev builds /sitrep skill to summarize multi-agent project state in one command — kevinkern · 2026-10-14
- Beancat launches an AI agent for media production with transparent per-run dollar pricing — Scobleizer · 2026-10-14
- Two devs burn $2,025 in API costs building a game with Claude, screenshots show progress — zeeg · 2026-10-14