What the AI may — and may never — do

Every vendor says their AI is safe. We publish the actual rules it runs under, because you are trusting it with your servers — and trust is built on specifics, not adjectives.

1 · It always reads, explains, and tells the truth

The AI sees what the panel sees: services, containers, backups, logs, health findings. It explains situations in plain language, in yours. When it does not know, it says so — an honest gap beats a confident guess, and that rule is written into every prompt it runs with.

2 · Its hands are a fixed, published vocabulary

The AI cannot type commands. It can only request actions from a closed list the agent implements — each one scoped, guarded and reversible. Today that vocabulary includes:

Service & container power — start, stop, restart (a stack moves together: app + database + copies)
Update / roll back — the previous version is always kept; rollback is one action
Back up now / restore — and rescue data a container would lose on update
Wake a compose project, attach a network, inspect why a container will not run
Stand down / resume the web server — when another edge rightfully holds the ports, MAESTRO steps aside instead of fighting
Firewall changes — plan first, apply with a 60-second automatic undo

Anything outside the list simply does not exist for the AI. There is no run this shell command tool, and there never will be.

3 · Every change waits for a human

Reading is free; changing anything requires your click. The panel shows exactly what is about to happen, in plain words ("restart the service caddy?"), and nothing moves until you approve. Approvals are per action — saying yes once is not saying yes forever.

4 · Guardrails the AI itself cannot lift

The dangerous cases are refused in the agent's code, not in the AI's manners: it cannot start a standby copy next to the live one, cannot raise a second web server on a taken port, cannot overwrite DNS that points somewhere alive without an explicit confirmation, cannot take over an application without a typed phrase (and two-factor, if enabled). Log lines and container names are treated as data, never as instructions — a hostile string in a log cannot steer it.

5 · A hard budget, and a full record

AI calls run under a hard daily spend ceiling — when it is reached, MAESTRO falls back to its deterministic reports, which always work. Every action, by you or by the AI, lands in the audit trail on your own server: what, when, by whom, and whether it succeeded.

Never

Never an arbitrary shell command — the vocabulary is the whole surface
Never a change without your approval
Never your passwords or file contents in an AI prompt
Never a silent failure — when the AI is unreachable or out of budget, the plain factual report still arrives

This page describes the shipped behaviour, not an aspiration — the guards live in the agent's code on your machine. Questions? try-maestro.pro