Ops AI
Ops AI is Panelica's most powerful feature and its most powerful weapon: a full desktop environment inside your browser, hosting real AI agents that administer the server for you. Not a chatbot, actual Claude Code running as terminal sessions on the server, backed by nineteen domain experts, a live browser they can drive, and MCP tools. You tell an expert what you want in plain language; it runs the commands.
Tools → Ops AIhttps://YOUR-SERVER-IP:8443/tools/ops-ai (ROOT only)● fail2ban active · 4 jails · 12 currently banned
203.0.113.9 (ssh) · 198.51.100.4 (panel) ...
What it actually is
Open Ops AI and you get a desktop: draggable, resizable windows, a taskbar, a start menu, four virtual workspaces (Ctrl+Alt+1 to 4), and a command palette (Ctrl+Shift+P or F1). Its apps are where the power lives:
- Experts (a folder of 19): each is a specialist AI agent in its own terminal window.
- Terminal: a plain root bash shell.
- Browser: a live Chromium running on the server that the AI can drive (see below).
- Prompts: a library of ready-made requests to copy.
- Tools: the catalog of MCP integrations.
- Settings and Update: preferences and one-click upgrade of the AI stack.
On first open, a boot screen offers a one-click install (nothing is installed until you ask); after that it goes straight to the desktop. Your window layout is remembered between sessions.
The 19 experts
Each expert is Claude Code pre-loaded with a specialist system prompt and Panelica's full API. Pick the one that matches the job:
| Expert | Ask it about |
|---|---|
| General | Anything; the all-rounder |
| Domain | Adding domains, subdomains, document roots, provisioning |
| Security | Firewall, fail2ban, ModSecurity, audits, incident response |
| Mailboxes, DKIM/SPF/DMARC, Postfix/Dovecot, deliverability | |
| Database | MySQL/PostgreSQL, users, grants, tuning, backups |
| WordPress | Installs, plugins, staging, hardening, malware cleanup |
| System | Services, processes, resource use, OS-level tasks |
| Docker | Containers, images, stacks, the template catalog |
| Backup | Backups, restores, schedules, remote storage, migration |
| Git | Repositories, deploys, pipelines, webhooks |
| Cloudflare | DNS, tokens, proxy, cache, origin certificates |
| FTP, SSH, Cron, Files, Migration, Performance, Incident, Monitoring | Their respective subsystems |
You choose the model (auto, Sonnet, Opus or Haiku), the effort level, and a permission mode (default, plan-only for read-only analysis, or accept-edits) per session. Past conversations are saved and resumable; a "continue latest" button picks up where you left off.
Examples: how you actually use it
The point is that you describe the outcome, not the commands. Real examples:
shop.example.com for user jane, install WordPress on it, and issue a Let's Encrypt certificate." The agent creates the domain, provisions WordPress, requests the certificate, and reports the admin URL and credentials.example.com is returning 502. Find out why and fix it." The agent checks the service status, the PHP-FPM pool, the nginx and Apache logs, spots the cause (say a crashed pool), restarts it, and confirms the site is back, showing you what it found.panelica.com in the browser, log into wp-admin with these credentials, and take a screenshot of the plugins page." The agent uses the Playwright browser tools to actually navigate, log in and capture the page, which you watch live in the Browser window.The live browser
The Browser app runs a real headless Chromium on the server (via Xvfb and noVNC through the panel's own port), which the experts drive with the Playwright MCP: they can navigate, fill forms, click, and screenshot. You watch it happen in the window, or click in it yourself. This is how the AI checks a site the way a human would, from the server's own network.
MCP tools
The experts come with the Playwright browser tools active by default. A catalog of other MCP integrations (filesystem, GitHub, git, Postgres, fetch, memory and more) can be added from the Tools app when you need them.
The honest security picture
This is the part to read twice, because Ops AI is genuinely powerful and you are trusting it accordingly.
public_html or the panel's own code). There is no per-command approval dialog. Treat every Ops AI session as if you had handed someone a root shell: use it on servers you own, review what an expert proposes on anything destructive, and prefer the read-only plan permission mode when you just want analysis. It is a superb operator and a serious responsibility.Access
A PRO and up licensed feature (ops_ai), and strictly ROOT at the menu, the route and every backend call. See roles and permissions. Nothing is installed by default; the first visit offers the one-click setup.