AI assistant — the one in the panel, and connecting Claude to your panel

Nimbopanel gives you two ways to use AI. Both are off until you switch them on, and in both the assistant can only see what the person using it may see.

The built-in assistant

People ask in their own language — "why is my email not arriving?", "how much disk is left?" — and the assistant reads the panel to answer.

Switch it on (administrator):

1. Open System → Security → AI assistant in the panel.
2. Choose the AI provider (see the list below). A link next to the choice takes you to where that provider makes keys.
3. Paste your API key for that provider. It is saved in a root-only file on your server and is never shown again — not to you, not to anyone. Switching provider does not carry the old provider's key over.
4. Choose the model. The list offers a few common names for the provider, and you can type any model name exactly as the provider lists it.
5. Choose who may use it: administrators, resellers, customers, team users.
6. Set the limits: questions per person per day and the monthly token budget for the whole server. When the budget is used up, the assistant pauses until next month.
7. Tick Switch the assistant on and save.

Providers you can choose. Ten are listed by name, plus two for any other address:

For the two Other choices you enter the provider's API address yourself, as its documentation gives it. The address must start with https://. Plain http:// is accepted only for this server or a private network (a model server you run yourself). An address with a user name, a query or a fragment in it, and addresses in the cloud metadata range, are refused.

Honest limit. The assistant has been tested with a stand-in provider that behaves like each protocol. It has not yet been tested against every vendor's live service, so try a question or two after you switch it on and before you tell your customers about it. Model names change often; if a provider answers that it does not know the model, type the current name.

Choosing an AI provider and keeping the bill small

• You pay the provider, by use — Cost grows with the number of questions and the length of the answers. Prices change often, so open the provider’s price page before you decide, and also set a spending limit in the provider’s own dashboard. Nimbopanel adds no charge.
• Which provider? — For good answers with the least work, choose a hosted provider such as Anthropic, OpenAI or Google. For one key that reaches many models, choose OpenRouter. To keep every question on your own server, run Ollama there: no per-question fee, but it needs enough memory and smaller models answer less well. You can change provider later.
• Start small — Keep the starting limits (30 questions per person a day, 2,000,000 tokens a month for the whole server) and let only administrators use it at first. 0 means no limit, so leave that for later.
• Decide what the AI may read — Tick “Keep support tickets away from the AI service” if you do not want customers’ messages sent to a provider. Every person sees a privacy notice before their first question.
• Try it before you tell customers — Ask a few questions on a test account, and try one proposed change: approving it asks for the password of the person approving.

What it can and cannot do

• It reads with the asking person's own access: a customer's assistant sees that customer's account and nothing else; a reseller's assistant keeps the reseller's privileges.
• It cannot change anything by itself. For a few changes (adding or deleting a DNS record, suspending or bringing back an account) it prepares a card under its answer. The card is written by the panel — not by the AI — and says exactly what will happen. Nothing happens until the person presses Approve… and then Yes, do it now. They can also press Reject. Because an assistant asked for the change, the panel first asks you to confirm it is you with your password (and your verification code if you use two-step sign-in); after that it does not ask again for a few minutes.
• When the assistant read messages written by other people (for example support tickets) before proposing something, the card says so. Approve only what you actually asked for.
• Requests wait 30 minutes, then expire. Switching the assistant off cancels every open proposal.
• It is not available in an API token or in a "Log in as" tab — only in a person's own session.

Privacy. Each person sees a notice before their first question: questions, and the panel data the assistant reads to answer them, are sent to the provider you chose, under your account. Passwords and keys are removed from what is sent where the panel can recognise them, but people should never type a password into the assistant. An administrator's or reseller's assistant may read other customers' data that they are allowed to see.

Connecting an AI assistant (MCP)

This lets someone use their own assistant (for example Claude) with the panel, through the Model Context Protocol.

1. Administrator: open System → Security → AI assistant connection (MCP), tick Allow AI assistants to connect, choose which roles may connect, save.

2. The person (administrator, reseller, customer or team user) opens API tokens → Connect an AI assistant (Claude and others). For each part of the panel choose No access, Read or Read and change — choose Read when the assistant should only look — then press Create assistant token. The token stops working after 90 days unless you set another number in Stops working after (days).

3. The panel shows the command to paste. For Claude Code:

   claude mcp add --transport http nimbopanel https://panel.yourdomain.com/mcp --header "Authorization: Bearer <token>"
   

   Other MCP clients use the same address with the token in the Authorization: Bearer header.

What a connected assistant can do: read the parts its token names; add an address or text record for a new, ordinary name. Everything else — deleting a record, suspending or bringing back an account, mail records, the domain itself, CNAME/SRV/CAA records, wildcards and names that mail programs look up (autodiscover, autoconfig, mail, …) — becomes a request. The token's owner sees it under API tokens → Assistant requests and approves or rejects it. An assistant can never approve its own request.

Not supported yet: clients that can only connect with OAuth sign-in (the endpoint accepts a token in the header). Revoke a token at any time under API tokens; its open requests are cancelled.

Good habits

• Give each assistant token only the parts it needs, and prefer Read over Read and change.
• Read every approval card. The card — not the assistant's message — says what will happen.
• Keep the daily question limit and the monthly budget set; check this month's use on the same card.