> ## Documentation Index
> Fetch the complete documentation index at: https://docs.vetox.io/llms.txt
> Use this file to discover all available pages before exploring further.

# AI Server Control

> An AI assistant that carries out server administration on request, with access control, protected resources and tool restrictions. Black tier.

<Info>
  **Black tier only.** On any lower tier the page is replaced by the plans view.
</Info>

Mention Vetox and describe the task in plain language, in any language. It plans the work, runs it, and reports what it did — across **182 actions in 27 categories**.

<Note>
  Two switches must both be on: **AI Server Control** in [Server Setup](/en/server-management/setup), and the settings on this page.
</Note>

## How you talk to it

Mention the bot directly. `@Vetox create a #announcements channel that only staff can post in`.

Three things are **not** triggers: `@everyone` and `@here`, a role the bot happens to hold, and the automatic ping on a reply.

<Tip>
  Once it has opened a thread, keep talking in that thread — you do not need to re-mention it. The conversation stays live for **two hours**.
</Tip>

### Thread or reply

| Mode                 | Behaviour                                      |
| -------------------- | ---------------------------------------------- |
| **Thread** (default) | Opens a thread on your message and works there |
| **Reply**            | Replies directly in the channel                |

In thread mode, **Private thread** is on by default, so the conversation is visible only to you and staff with access. If a private thread cannot be created, it falls back to a public one.

## Access — read this before enabling

<Warning>
  **Access is deny-by-default.** A member who is neither an Administrator nor explicitly allowed can never use the AI, no matter how the switches are set.
</Warning>

The Access tab shows an "Effective access" summary. When it reads *"No restrictions set — anyone can use the AI"*, that line is misleading — the real behaviour is still Administrator-only.

Two switches decide who gets in:

| Require Administrator | Allowlist bypasses Administrator | Who can use it                                      |
| --------------------- | -------------------------------- | --------------------------------------------------- |
| **On** (default)      | **Off** (default)                | Administrators only. **The allowlists do nothing.** |
| On                    | On                               | Administrators, plus anyone on an allowlist         |
| Off                   | either                           | Administrators, plus anyone on an allowlist         |

<Tip>
  To let a trusted moderator use the AI without giving them Administrator, add them to **Allowed users** and turn **Allowlist bypasses Administrator** on. Adding them to the list alone has no effect.
</Tip>

### Per-tool permission is a second gate

Being allowed to *talk* to the AI does not grant it your Discord permissions. Every action also re-checks the matching permission on the person who asked.

**Bypass roles** and **Bypass users** skip those per-action checks. Grant this carefully — it lets someone perform actions through the AI that they could not perform themselves.

## Tools

27 categories covering channels, roles, moderation, messaging, threads, voice, webhooks, invites, emojis, stickers, events, forums, onboarding, polls, integrations, soundboard, templates, the widget, the welcome screen, the vanity URL, and more.

Switch a whole category off and every action inside it becomes unavailable, even when explicitly asked for. Expand a category to switch individual actions.

<Note>
  The category switch wins. With a category off, its individual switches are greyed out and irrelevant. Turning the category back on restores whatever per-action state you had saved.
</Note>

Actions that cannot be undone are marked **Destructive**.

<Tip>
  Use the search box at the top — with 182 actions, finding "ban" or "webhook" is much faster than scrolling.
</Tip>

### Confirmation is not optional

Every destructive action is held for your confirmation before it runs. There is no setting to disable this.

The confirmation card lists exactly what is queued. Only the person who made the request can confirm it, and it expires after one hour. On confirmation, **every access check is re-run against the live configuration** — so revoking someone's access invalidates their pending confirmations too.

## Scope

**Where the AI works** — all channels, only selected channels, or all except selected. Threads inherit their parent channel's scope.

**Protected roles and channels** — the AI refuses any action targeting these: no editing, deleting, cloning, moving, assigning or posting.

<Warning>
  Protection is absolute. A protected target cannot be touched **even with admin confirmation**. This is the correct place to lock down your Admin role, your rules channel and your log channels.
</Warning>

<Tip>
  Protected channels can include categories, so protecting one category protects everything filed under it.
</Tip>

## Persona

Up to 2,000 characters setting the tone and any standing instructions — *"Be concise and formal. Always confirm before deleting channels. Refer to staff as the team."* Leave it blank for a neutral default.

## What a run looks like

<Steps>
  <Step title="It starts working">
    A typing indicator appears. Simple questions are answered directly without the full toolset being loaded.
  </Step>

  <Step title="A progress card appears">
    Created only once it actually changes something — a pure lookup never shows one. Each action adds a line: ✅ done, ❌ failed with the reason, ⏳ waiting for confirmation.
  </Step>

  <Step title="Confirmation, if needed">
    Destructive actions surface the confirmation card. Nothing in it has run yet.
  </Step>

  <Step title="It reports back">
    A green summary of what changed, or a plain answer if you only asked a question.
  </Step>
</Steps>

<Note>
  A single request runs up to **8 steps** within a **90-second** budget. Very large requests may stop early and say so — break them into smaller ones.
</Note>

## Rate limits and concurrency

* Roughly **one request every 5 seconds** per person, with a small burst allowance
* **One request at a time** per person
* **Four at a time** across the whole server

Hitting the per-person or per-server limit produces a short "still working / try again in a moment" notice, at most once every 20 seconds.

## Budget

A run needs at least **30,000 tokens** in reserve. Below that it declines with a budget-exhausted message rather than starting work it cannot finish.

<Card title="AI Usage" icon="chart-line" horizontal href="/en/moderation/ai-usage">
  Track spend, see which actions ran, and top up.
</Card>

## Troubleshooting

<AccordionGroup>
  <Accordion title="The bot ignores me completely">
    Silence is the normal response to every access failure — there is no error message by design. Check in order:

    1. Did you **mention the bot directly**? A role ping or a reply ping is not a trigger.
    2. Is **AI Server Control** on in Server Setup?
    3. Is the channel in scope?
    4. Are you an Administrator, or on an allowlist **with Allowlist bypasses Administrator turned on**?
    5. Have you sent several requests in a row? Wait a few seconds.
  </Accordion>

  <Accordion title="I added someone to Allowed users and it changed nothing">
    Expected. With the defaults, allowlists are inert. Turn on **Allowlist bypasses Administrator**, or turn off **Require Administrator**.
  </Accordion>

  <Accordion title="It refuses an action it clearly has a tool for">
    Either the tool's category is switched off on the Tools tab, or the target is on your protected roles or channels list, or the requester lacks the underlying Discord permission for that action.
  </Accordion>

  <Accordion title="It stopped partway through a big job">
    Eight steps and ninety seconds per request. Split the work — "create these five channels" then "now set permissions on them".
  </Accordion>

  <Accordion title="It says the budget is used up">
    Below 30,000 tokens it will not start. Top up on the AI Usage page or wait for the monthly grant.
  </Accordion>

  <Accordion title="A confirmation expired">
    Confirmations last one hour. Ask again.
  </Accordion>
</AccordionGroup>

<CardGroup cols={2}>
  <Card title="AI Auto-Moderation" icon="sparkles" href="/en/moderation/ai-auto-mod">
    Context-aware content moderation.
  </Card>

  <Card title="Access Control" icon="key" href="/en/server-management/dashboard-permissions">
    Who can change what on the dashboard.
  </Card>
</CardGroup>