OpsAI: running Claude Code, Codex CLI, Gemini CLI and Cursor Agent inside the panel

[admin]

OpsAI: an AI desktop built into the panel​

OpsAI is a windowed AI desktop that lives inside the Panelica panel itself: a terminal running a coding agent as an isolated system user, a live browser window the agent can drive, a Telegram bridge so you can reach it from your phone, and 19 built-in experts — all without opening a terminal on your own machine or installing anything locally. It's a different surface from the Panelica MCP Server covered in the complete guide linked at the bottom of this thread: OpsAI runs the agent inside the panel and gives it a desktop; the Panelica MCP Server runs outside the panel (on your own laptop, wired into your own Claude/Cursor/Codex) and gives your assistant tools to call the panel's API. OpsAI is ROOT-only and sits behind the

ops_ai

license feature.

This thread covers what OpsAI actually ships today: the four providers you can install, exactly how login and permission modes work for each, the 19 built-in experts and what they're for, the live browser window, the Telegram bridge, who can access any of it, and where the honest limits are.

Why run an agent inside the panel at all?​

The Panelica MCP Server is the right tool when you already live in Claude Code, Cursor, or Codex CLI on your own machine and want your assistant to reach into the panel as one more tool among the others it already has. OpsAI is the right tool for the opposite situation: you're already in the panel, you don't want to context-switch to a local terminal, install a CLI on your laptop, or manage an API key yourself, and you'd rather the assistant's terminal, browser and chat all live in the same browser tab as the rest of the panel. It also gives you a Telegram bridge for the moments you're not at a desk at all. Neither replaces the other — they're two different points of entry to the same panel.

How is OpsAI different from the Panelica MCP Server?​

Both let an AI agent operate a Panelica panel, but they run in different places and suit different situations:

	OpsAI	Panelica MCP Server
Where it runs	Inside the panel, in a windowed desktop	Outside the panel — on your laptop, in your own MCP client
Who talks to it	You, through the panel UI (or Telegram)	Your own Claude Code / Claude Desktop / Cursor / Codex CLI
Access control	ROOT only, ops_ai license feature	Whatever scopes you put on the API key you generate
Can drive a live browser	Yes — Xvfb + noVNC window via Playwright MCP	No — it calls the panel's HTTP API directly
Mobile access	Yes — Telegram bridge	Only via whatever your MCP client offers
Providers available	Claude Code, Codex CLI, Gemini CLI, Cursor Agent — pick one or several	Any MCP-compatible client

If you'd rather run the assistant on your own machine with your own subscription and your own client, the Panelica MCP Server is the one to read about — see "Panelica MCP Server — complete guide" in the API & Automation section. This thread is about the panel's own built-in option.

Which AI providers can I run inside OpsAI?​

OpsAI supports four providers. You install whichever you want from a picker and set one as default; every provider gets the same window and the same 19 built-in experts, so the choice of provider doesn't change what you can ask it to do — only which company's model answers.

Claude Code (npm

@anthropic-ai/claude-code

) is the provider we have the most real, hands-on experience with inside OpsAI. Login is interactive. It exposes three permission modes —

default

,

acceptEdits

(skip confirmation on file edits) and

plan

(plan first, act only after you approve) — three model choices (

sonnet

,

opus

,

haiku

), and an effort setting for how much reasoning it does before answering.

Codex CLI (npm

@openai/codex

), OpenAI's coding agent, logs in through a device-code flow — you're given a URL and a short code to enter elsewhere, since the server itself has no browser to log in from — or you can paste an API key directly instead. It runs in

full

or

readOnly

mode.

Gemini CLI (npm

@google/gemini-cli

), Google's coding agent, logs in through an interactive auth dialog or an API key, and also runs in

full

or

readOnly

mode.

Cursor Agent, installed via the official cursor.com install script (the resulting binary is called

agent

), logs in interactively or via an API key, and runs in a single

full

mode.

How do I install a provider?​

Installation is a background job with a live step list — Node runtime, CLI install, Playwright MCP registration, verification — so you can watch it progress instead of staring at a blank screen while it works. The picker shows which providers are already installed and which are signed in, so you can tell at a glance whether a provider needs an install, a login, or is ready to use. Each provider gets its own context file (

CLAUDE.md

for Claude Code,

AGENTS.md

for Codex CLI,

GEMINI.md

for Gemini CLI) preloaded with the panel's own playbook — the same kind of grounding the Panelica MCP Server gives an external client through its server instructions, just delivered as a context file instead of an MCP

initialize

payload. Each provider also gets its own isolated working directory per purpose, so switching between, say, Claude Code and Codex CLI doesn't mix up one provider's working state with another's.

What are the 19 built-in experts?​

Every provider gets access to the same set of purpose-built experts — pick one to scope the conversation to a specific part of the panel instead of starting from a blank "General Assistant":

Expert	Purpose
General Assistant	General server management assistant
Domain Expert	DNS, SSL certificates, virtual hosts
Security Expert	Firewall, malware scanning, WAF
Email Expert	Postfix, Dovecot, DKIM/SPF
Database Expert	PostgreSQL, MySQL optimization
WordPress Expert	WP-CLI, plugins, themes, cache
System Expert	Services, disk, CPU, backups
Docker Expert	Containers, images, volumes, compose
Backup Expert	Backup, restore, snapshots, schedules
Git Expert	Repositories, branches, deployments
Cloudflare Expert	DNS proxy, cache, SSL, zones
FTP Expert	FTP users, quotas, transfers
SSH Expert	SSH users, keys, configuration
Cron Jobs Expert	Scheduled tasks and automation
Migration Expert	Panel-to-panel migration, cPanel/Plesk/DA imports
Performance Expert	Tuning, caching, slow requests, resource hogs
Incident Responder	Outages, 502s, crash triage, restart order
Monitoring Expert	Metrics, alerts, resource trends
File Manager Expert	Files, permissions, disk usage

Each one ships with its own quick-prompt suggestions, so opening, say, the Incident Responder gives you one-click prompts instead of a blank chat box. A sample of what those look like:

Domain Expert

• "Check DNS records for a domain"
• "Diagnose SSL certificate issues"
• "Review nginx vhost configuration"
• "List all hosted domains"

Security Expert

• "Run a security audit"
• "Check fail2ban status and recent bans"
• "Review ModSecurity events"
• "Analyze firewall rules"

Database Expert

• "Check database performance"
• "Find slow queries"
• "Analyze database sizes"
• "Review connection pools"

Incident Responder

• "A site returns 502 — find the cause and fix it"
• "Show errors from all services in the last 10 minutes"
• "Check whether the panel backend is healthy"
• "Restart customer services in the correct order"

Backup Expert

• "Show backup status and schedules"
• "Create a full server backup"
• "Check backup chain health"
• "List available snapshots"

Performance Expert

• "Find the top CPU and memory consumers right now"
• "Check PHP-FPM pool saturation per site"
• "Review MySQL slow query log"
• "Suggest safe nginx/PHP tuning for this server"

Migration Expert

• "List migration jobs and their status"
• "Run a migration health-check for a source server"
• "Explain what a migration will change before running it"
• "Verify a completed migration (files, DB, email)"

Cloudflare Expert

• "List all Cloudflare zones"
• "Purge cache for a domain"
• "Sync DNS records with Cloudflare"
• "Check Cloudflare connection status"

These are the exact suggestions built into the panel — every expert has four of them, and they're there to give you a starting point, not a limit on what you can type. Switching experts mid-conversation is normal — start in the Domain Expert to sort out a vhost, then hand the same terminal over to the Security Expert if the trail leads to a firewall rule, without leaving the OpsAI window at all.

Can OpsAI control a browser?​

Yes. Alongside the terminal, OpsAI gives the agent a live browser window — an Xvfb display with a noVNC viewer you can watch in real time — that it drives through the Playwright MCP. That means you can watch the agent open a page, click through it, and see exactly what it sees, rather than trusting a text description of what it "should have" done. This is the same Playwright MCP registration step you see listed during a provider's background install.

Can I use OpsAI from my phone?​

Yes, through the Telegram bridge. OpsAI runs on your desktop (inside the panel) and connects to a Telegram bot; any message you send in the configured chat gets picked up and answered by whichever provider you've chosen, in Telegram-formatted replies, and longer-running tasks keep running in the background — up to 30 minutes — while you go on using Telegram normally. There's a safe mode that forces the session into plan/read-only behavior for anything you trigger from Telegram, so a message you didn't mean as a command can't make a destructive change from your pocket, and bot tokens are redacted from logs and error messages so they don't end up on screen or in a support ticket. Full setup steps — messaging @BotFather, getting a bot token, creating the group chat — are posted in "OpsAI on Telegram: your server now answers your messages" in Announcements.

Can I look back at past OpsAI sessions?​

Yes. ROOT can pull up a history of past sessions — deduplicated by session so you see one entry per conversation rather than one per message, sorted newest first, capped at the 100 most recent — and open any one of them to see its full output log. Useful for checking what an unattended Telegram-triggered session actually did overnight, or for re-reading how an expert diagnosed something last week. That history is Claude Code's own session log specifically, read from its isolated working directory rather than reconstructed after the fact, so what you see in the detail view is what actually happened in that terminal.

Who can access OpsAI, and is it safe?​

OpsAI is ROOT only — it isn't something you hand to a reseller or a customer account — and it's gated behind the

ops_ai

license feature, so it has to be explicitly enabled on your license before the desktop even appears. The agent itself runs as an isolated system user (

claude-runner

), separate from the panel's own service account, rather than executing arbitrary shell commands as root directly. Combined, that means: only your root operator can open OpsAI, the license has to allow it in the first place, and the process doing the work is not the same account the panel itself runs as.

If a provider doesn't respond, what should I check?​

• Open the provider picker and confirm the provider shows as both installed and signed in — a provider that's installed but not signed in will not respond until you complete its login step.
• If you just installed it, check the background install's step list finished all steps (Node runtime, CLI install, Playwright MCP registration, verification) rather than stopping partway.
• For the Telegram bridge specifically, re-check the bot token and group chat ID — a wrong or revoked token is the most common reason messages go unanswered.
• If the live browser window won't load, that's the Playwright MCP / Xvfb+noVNC stack — try reopening the OpsAI window, which restarts that stack.

What are the honest limits?​

We have real, hands-on experience with Claude Code inside OpsAI. We have not benchmarked the chat quality of the Gemini CLI or Cursor Agent providers — we don't have accounts to run extended sessions with them the way we have with Claude. Both are real, supported providers with the login/mode details listed above, and both get the same window, browser, Telegram bridge and 19 experts as Claude Code — we're just not going to claim specific results for them here that we haven't measured. If you've run Gemini CLI or Cursor Agent inside OpsAI for a real task, we'd like to hear how it went — post below.

Frequently asked questions​

Do I need four separate subscriptions to use OpsAI? No — install and sign into whichever single provider you already have access to. The picker lets you add more later; you don't need all four.

Is OpsAI the same product as the Panelica MCP Server? No. OpsAI is a desktop inside the panel that runs a coding agent for you. The Panelica MCP Server is a separate, open-source package you run on your own machine that lets your own Claude/Cursor/Codex call the panel's API. See the comparison table above.

Can a reseller or a regular user open OpsAI? No — it's ROOT only.

Does OpsAI need internet access from the server? It needs outbound access for the CLI installs, provider login/auth, and, if you use it, the Telegram bridge.

Which provider should I pick first? Whichever one you already have an active account or API key for — all four get the same window, browser and experts, so the choice mostly comes down to which subscription you're already paying for.

Can I switch providers later? Yes — the picker supports installing more than one and switching the default at any time.

What does "safe mode" do on Telegram exactly? It forces the session into a plan/read-only permission mode for anything triggered from a Telegram message, so a message you didn't intend as a command can't make a destructive change.

Does the browser window see real traffic, or a sandboxed copy? It's a live browser window on the server, driven through the Playwright MCP — what you see in the noVNC viewer is what the agent is actually doing.

Where do the 19 experts' definitions live? They're built into OpsAI's backend as a fixed list, each with its own quick-prompt suggestions; you don't create or edit them per panel today.

Is there a way to try panel AI features without an AI subscription of my own? Not for OpsAI — you need a login or API key for at least one of the four providers. If you want to try the panel's AI integration without a provider account on the server itself, ask a colleague who already has Claude/Codex/Gemini/Cursor access to add the Panelica MCP Server (covered in the complete guide) to their own client instead — that path needs only a scoped API key from your panel, no provider install on the server at all.

Where to go next​

• Panelica MCP Server — complete guide (setup for every client, scopes, examples, troubleshooting): posted in API & Automation
• Announcement: "Panelica MCP Server 0.5.1: connect Claude, Cursor, Codex and Gemini to your panel" in Announcements
• "OpsAI on Telegram: your server now answers your messages" in Announcements — full Telegram bridge setup
• "Which permissions should you give an AI assistant? (scope cheat-sheet)" in this section

As of September 2026, OpsAI ships four providers and 19 built-in experts. The Panelica MCP Server — the outside-the-panel option, for anyone who wants to drive the same panel from their own Claude, Cursor or Codex install — is at version 0.5.1: 404 tools, 51 categories, 37 registered by default.

Questions and real-world experience reports — especially with Gemini CLI or Cursor Agent — welcome below.