Run a Discord AI Ticket Bot on Your Own API Key

Every AI ticket bot sells you inference. You pay a monthly price, you get an allowance, and when the allowance runs out the AI stops. That is a good deal until your server is large enough that the allowance is the only line on the invoice you think about.

What does connecting your own AI actually do?

It moves one thing: who runs the model call.

Everything else stays where it was. Your server's brain, the facts your staff taught it, the rules you set, your panels, your transcripts and your staff permissions are all ours and all untouched. Connect a key and the next reply is generated by your provider instead of by ours. That is the whole change.

The scope of "everything" is worth stating, because people assume it means ticket replies only.

That Manage Server requirement is stricter than most of the dashboard on purpose. A provider address receives every conversation from the moment it is connected, so this is a data decision rather than a settings change, and it belongs with the people who already control the server.

Why would you want to?

Because on a normal plan the token allowance is the ceiling, and your own key replaces it with your own.

Pro includes 15,000,000 AI tokens a month. An average ticket conversation runs about 13,000 billable tokens across roughly five turns, so that is somewhere near 1,150 conversations before the AI stops cleanly and your staff carry the rest by hand. For most servers that is far more headroom than they will ever use. For a few, it is the number they hit on the 20th.

That last row is the real trade and it cuts both ways. A fixed price suits a server that wants one predictable invoice; usage billing suits one whose volume swings between a quiet week and a launch. If you sit comfortably inside the allowance, the plan you already pay for is the cheaper answer.

What still comes out of your plan?

Web searches, and only web searches.

When the AI cannot answer from what your staff taught it, it can check the live web as a fallback. That lookup runs on our side whichever AI is generating the reply, so it is charged to your plan at roughly 5,000 tokens a search. Everything else on that turn is on your provider's bill.

Searching the AI's own memory is not charged to you either: that runs on your provider like any other model call. If you want the plan meters to stop moving altogether, web search has a server-wide switch in /ai settings.

The arithmetic is worth doing once. If your plan is only paying for searches, Pro's 15,000,000 tokens is in the region of 3,000 web searches a month, which is not a number a support queue reaches.

Which providers does it work with?

Nine have a preset, and the tenth option covers everything else.

You never have to know which API format your provider speaks. The address is checked once, the format is worked out from the provider's own answer, and the key is tested with one real call before anything is saved.

Two limits sit here, both worth knowing before you buy credit. A model that cannot use tools is refused. The AI needs them to look something up, close a ticket or write to memory, so the test call fails rather than the feature quietly degrading. And our prompts are written and tuned against Claude, so another model will work and may read a little differently. We tested a spread of models before shipping this: most handled every task, a few invented an answer or skipped a tool entirely. Pick a current model rather than the cheapest on the list.

What happens when the key stops working?

This is the part that decides whether the feature is safe to leave running unattended, and nobody asks about it until it happens.

Four failures are treated as things only an admin can fix: the key is rejected, the account is out of credit, the model is gone, or the provider will not run the model in a way the AI can use. Any of those and the server moves back onto our AI immediately, the member's message is sent again, and they get their answer on that same turn rather than the next one. Whoever connected the key gets one direct message, once per incident, and the AI page shows what the provider actually said.

Anything that looks temporary is treated as temporary. A rate limit, a slow provider, a provider that is briefly down: none of those move your server off its own AI, because a short outage is not a reason to switch a whole server onto a different model.

Coming back is deliberate rather than automatic. Fix the problem, then press Test and switch back, or save a new key or model that passes its test. A key that has already failed is never retried against a member's question on its own.

What does not change

The list of things people expect to lose here is longer than the list of things that actually move.

The brain stays. Memory, rules, the facts your staff taught with /ai train and every past ticket live on our side and are not touched by connecting, switching or removing a key. What the AI remembers about your server covers where that sits and who can edit it.

The key is held once and never shown again. It goes to our AI service, is encrypted there, and after that only its last four characters appear anywhere. It is never stored on the website, never written to a log and never included in an error message.

A plan lapse pauses instead of deleting. Drop below Pro and the switch pauses with the key kept, then comes back by itself when you renew. Renewing restores only what the plan paused, so a switch you turned off yourself stays off.

Where this is the wrong choice

The cheapest-answer line is the one worth arguing about. A Discord ticket bot is judged on whether the answer was right, and money saved on a cheaper model comes back as staff time spent correcting it. If the reason for connecting a key is headroom, or a data agreement, this is a good feature. If the reason is shaving the cost of a reply, price the staff hours too before you switch.

One practical note for whoever owns the account: the failure that actually happens is not a broken integration, it is a card that expired on an account nobody logs into. Put the provider balance on the same list as your domain renewal. The bot handles that failure gracefully and still cannot pay the bill for you.