AI
Elpan is built to be operated by an AI as well as by a person. The AI isn't a help chatbot beside the panel: every panel feature is an operation (create a database, read a file, replace a DNS record, restart a container), and an AI calls those operations directly, with the same validation, permissions, and audit trail as the buttons in the interface.
There are two ways to use this. The one real difference is whose AI does the work.
| AI Assistant | MCP server | |
|---|---|---|
| Where it runs | A chat inside the panel. | Your own AI tools: Claude, an IDE assistant, or any Model Context Protocol client. |
| Who supplies the model | Your provider, from a budget they grant you. | You. |
| What limits it | You approve each write. | The scopes you approve at consent, and your client's confirmation policy. |
Both reach the same operations, and in both cases the AI stays inside your account: it authenticates as you, can do nothing your own user can't, and every action is recorded.
What it's useful for
An AI that can act, not only answer, closes the distance between diagnosis and repair:
- Debugging what's broken. "My site returns a 500": a model can read your error logs, open the file the trace points at, and check which PHP version the site actually runs, reaching the cause in one pass. See Logs.
- Work that spans several screens. Putting an application online touches a database, an app server, a mount, and DNS. Describe it once in a sentence; the AI does it as a sequence and shows you each write first.
- Bulk and repetitive changes. Address changes across a domain, a batch of DNS records, file cleanup after a migration: slow in a form, short to describe.
- Questions about your own setup. Not "how does DNS work" but "which of my domains has no SPF record". Only something with access to your account can answer that.
Why it stays safe
Both ways rest on the same foundation:
- The AI acts as you, never above you. It authenticates on the server as your account and runs commands as your own system user. There's no elevated path: what your user can't do, the AI can't.
- A write needs approval. The assistant stops at each change and waits for Approve or Reject. MCP access is limited by the scopes you approve, and your client applies its own confirmation policy.
- A read is free. Listing files, reading logs, and inspecting configuration only look, so they need no approval. This is what makes diagnosis fast.
- Everything is recorded. Each command and operation goes into your audit trail, whether from the panel, the assistant, or a connected MCP client.
"The AI acts as you" cuts both ways. A destructive command you approve is still destructive.
In this section
- AI Assistant: the built-in chat, what it can do, the approval model, and the usage budget.
- MCP server: connecting your own AI tools, the OAuth consent, the scopes, and what the tools cover.
If your code decides what to call and no AI is involved, the same operations are available as a conventional REST API with API keys.