We have written about what OpsAI is and what Sentinel does. The question we keep getting in tickets is narrower and more useful: what does a real job look like, from the prompt to the result? So here are six, written the way they actually happened — the request, what the agent did, and what you should check afterwards.
Two things to keep in mind while reading. First, OpsAI drives your own agent: Claude Code, Codex CLI, Gemini CLI or Cursor Agent, installed from the panel and signed in with your account. We do not run a model for you. Second, the agent is not poking at config files blindly — it is given the panel's own API, with an index of 1388 endpoints across 93 categories, so it changes things the same way the interface does.
Job 1 — "example.com returns 502. Find out why and fix it."
This is the single most common request, and it is a good illustration of why an agent with API access beats an agent with a shell.
- It reads the site's error log and the PHP-FPM log for that account — through the panel, so it looks in the right place for that domain rather than guessing a path.
- It finds the pool refusing connections because it reached its process limit.
- It raises the limit for that pool and reloads PHP-FPM.
- It reports what it changed, with the log lines that led it there.
What to check afterwards: the new pool value in Domains → PHP Settings, and whether the limit was the cause or a symptom. If the pool filled because one script is slow, raising the ceiling buys time; the agent will usually say so.
Job 2 — "Move this account from our old cPanel server."
Pick the Migration expert first. The expert is not decoration: it appends a role brief to the agent's system prompt, so the agent follows Panelica's migration pipeline instead of inventing one.
The run goes through discovery (what is on the source), credentials, files, databases, mail, SSL, and a verification pass. Database user password hashes are transferred as hashes, so applications keep working without anyone resetting passwords.
What to check afterwards: the verification output, then the site in a browser. If DNS still points at the old server, the agent will tell you that too rather than claim success.
Job 3 — "Why has the server been slow since this morning?"
With the Performance expert, this stops being a guessing game. The agent looks at load, I/O wait, slow queries and the busiest accounts, and comes back with a ranked list: which account, which process, which query. The useful part is the ranking — "these three things changed, this one explains the load" — rather than a wall of statistics.
Job 4 — from your phone: "disk is filling up, what can I safely remove?"
OpsAI has a Telegram bridge. You message your own bot, the agent answers there, and the conversation resumes where it left off. It replies with the largest directories, what is safe to delete and why — backups still inside their retention window, for instance, are reported as not safe.
This is the part people underestimate: the useful moment for server work is rarely when you are sitting at a desk.
Job 5 — a page that only misbehaves in a browser
Some problems do not appear in logs: a checkout that fails on the third step, a redirect loop that only triggers with a cookie set. OpsAI ships a real Chromium on the server, shown inside the panel. The agent can open the page, click through, and read what happens. For "it works with curl but not in a browser", this is the difference between a guess and an answer.
Job 6 — the one you do not ask for: Sentinel
Sentinel is the same agent, but it is not sitting in a chat window waiting for you. It is the guard on the dashboard's bottom bar, and its rule is simple: the AI never polls. The panel watches — 40 triggers across panel events, thresholds, log patterns and five-minute probes — and an agent run happens only when one of them fires. A quiet server costs you nothing.
Two runs from our own test server, with the timings as we measured them:
An [emerg] line in nginx's error log
- Queued 2 seconds after the line was written.
- Diagnosed and closed 23 seconds later.
- The agent noticed the line was an injected test and lowered the severity itself — which is exactly the behaviour you want from something allowed to wake you at 3 a.m.
A probe found 12 failed systemd units
- 48 seconds later there was a report waiting, with a proposed action.
- It was waiting, not done, because the level was inform.
- Steps outside the allow-list were listed as suggestions you can read — and if you agree, one button hands exactly those steps back to the agent for a single authorised run.
How much you let it do
Three levels, and the limit is enforced by the agent's own command line, not by asking it politely in a prompt:
- inform — read-only diagnosis. It writes a report; anything it wants to do waits for your approval.
- safe — read-only diagnosis, and it may request actions from a fixed allow-list, which the panel executes and then verifies.
- full — the agent acts itself.
The allow-list is deliberately short: restart a service, reload nginx, Apache or PHP, clean temporary files, vacuum the journal. Everything else becomes a suggestion. On top of that there is a cool-down per trigger, a window counter, a loop breaker so nothing can be restarted in circles, quiet hours that critical events still pass, and a daily cap. Cool-downs survive a panel restart.
What it costs
You pay your agent provider, and only for runs that actually happen. There is no Panelica subscription for the model, no per-token markup, and no background polling inflating the bill. A server with nothing wrong produces no runs.
Getting started
Tools → Ops AI, pick an agent, press Install — it runs as a background job with a live step list, because on a clean server it takes a few minutes — then sign in. Sentinel is the first tab of the bar at the bottom of the dashboard. It starts switched off and asks you to choose a level before it does anything.
If you want a setup for a specific case — which triggers, which level, which expert — open a ticket or ask on the forum and we will answer with the concrete configuration rather than generalities.