Assistant settings¶
Managers set the assistant up in Settings, Assistant. It is off until you turn it on. Changes are kept with the page's Save changes.
Turned off, nobody can ask it anything. Its users keep their conversations, to read or delete them until they expire, 30 days after their last activity; the button stays for them meanwhile.
Provider and key¶
The assistant uses your own key of an AI provider: Anthropic, or OpenAI, Google Gemini and OpenRouter (marked beta: not yet tested with the panel's tools). The panel talks to each provider at its own fixed address and to no other: there is no field for a web address. A self-hosted panel needs outbound access to the provider's address (and to GitHub's, for code).
Paste the key under Provider key and save: the panel lists the provider's models with it first, so a wrong key, or a panel that cannot reach the provider (a self-hosted panel without internet access), is told at once. The key is stored encrypted and never shown again; type a new one to replace it, or press Remove the key (the assistant goes off).
Every user the assistant serves uses this one key, each with their own role.
Changing the provider. A key belongs to one provider, so type the new provider's key in the same save. The models are the new provider's: the allowed models and the default one start empty and the assistant goes off: Assistant on cannot be ticked until the new key is saved. Choose the models, then switch it on as the first time, box included, and save. Everyone sees the first-use notice again, now naming the new provider.
Switching it on¶
Assistant on needs a key, at least one allowed model and a default model, and a box ticked with it. The box is asked each time the assistant is switched on, and not again while it stays on. The panel records the date, who ticked it and from which address, and shows it under the form.
IncubaCloud SaaS
The box is I accept section 12A (AI features) of the Terms of Service, with a link to the Terms: they say what is sent to the AI provider, who pays for it, and which actions wait for confirmation. IncubaCloud records your acceptance on your account first, with the version of the Terms, and the panel shows it: Section 12A of the Terms accepted on … by … from … (terms of …). If it could not be recorded, the assistant stays off and the panel says so; try again later.
On a self-hosted panel there are no IncubaCloud Terms to accept. The box is I understand that what the assistant reads is sent to my AI provider, and that it can make mistakes, with no link. It is recorded in the panel only, without a version: Last switched on, with the box ticked, on … by … from ….
Models¶
The list comes from the provider, with the key; Update the list reads it again. Choose the models your users may pick (Allowed), and the default one (Default). Recommended models were tested with the panel's tools; Untested ones may fail, mostly at writing modules.
Price unknown: the panel does not know that model's price. Without a monthly limit it can be used, and only its tokens are counted. With a limit it cannot: its cost could not be counted against the limit, so you cannot allow it (Not usable with a monthly limit), users see it disabled in the chat, and the default model must have a known price.
The panel asks the provider for the model chosen and no other: it never lets the provider fall back to another model, and an answer from a different one is set aside and recorded in the audit log. A conversation whose model is no longer allowed moves to the default one at its next message, and says so.
Users¶
Only the users you tick see the assistant. Choose them one by one; their role still decides what the assistant may do for them.
What asks for a confirmation is not a setting: the same rules hold on every panel (see the assistant). After the assistant reads text other people can write — logs, repository files, mail, a command's output — every change of that task asks the user, the work on its own new branch included, until the user writes again.
Limits¶
- Monthly limit per user (EUR): one value, the same for every user: what each user's conversations may cost per month, in euros, approximately (0: no limit). When a user reaches it, the assistant stops for them until next month or until you raise it. Each user sees their own spending in the chat.
- Steps per task: calls to the model one task may make before it stops and asks whether to go on (30 by default, from 1 to 200).
Usage¶
The table shows each user's tokens and approximate cost per month. Costs are estimates from the provider's public prices, in euros; your provider's invoice is what you pay. A row marked with * includes calls of models whose price is not known: only their tokens count.