Opening Pandora's Box Carefully

Hi, Paul here yet again (must be sick of me by now, I should make someone else write something), CEO of CreeperHost.

We have talked about the hardware. We have talked about the network. We have talked about the systems behind the scenes, and the panel you actually use.

Now it is time to talk about the new bit inside CreeperPanel.

Theo.

If you have been using our prerelease panel, you may have seen Ask Theo starting to appear. At first our AI work was narrow and mostly hidden away. Crash report analysis. Task scheduler help. Support context. Small tools to help turn messy technical information into something people could actually use.

That was useful, but it was not the same as putting a full assistant into the panel.

So as Theo becomes more visible, I want to be very clear about what we are building, what we are not building, and why we are being careful about it.

I Am Not Completely Comfortable With AI

That is probably not the normal way to start a blog post about launching an AI assistant.

But it is true.

I am not completely comfortable with AI. I do not trust it by default, and I do not think anyone should.

That does not mean I think we should ignore it. CreeperHost has always been early to look at new technology, especially when it might help players, server owners, modders, and our own team solve real problems. Sometimes that has meant building our own hardware. Sometimes it has meant writing our own systems. Sometimes it has meant doing things the awkward way because the easy way would mean handing control to someone else.

AI is another one of those moments.

Pandora's box is open now. Pretending otherwise will not protect anyone. The real question is whether we can build with it honestly, carefully, and without forgetting the people around it.

The thought behind our whole approach is simple:

Fear should never stop progress, but progress should never harm people.

That is the line.

Theo Is A Tool, Not A Replacement

CreeperHost is not a company full of placeholders waiting to be automated away.

Our team has people who have been around games, hosting, Minecraft, modding, modloaders, MMO server reverse engineering, infrastructure, support, and weird edge cases for a long time. Not "I read a blog post about it" long. Actually there, fixing it when it breaks, long.

That matters.

Some problems are not hard because nobody can read the log file. They are hard because the log file is only one piece of the story. The real answer might be in the modpack, the JVM flags, the panel state, the network path, the save file, the update that ran an hour ago, or the thing the customer did not know was important enough to mention.

Theo can help gather that context.

He can explain a crash report. He can look at service state. He can help prepare certain panel actions. He can help a customer at 2am when the server will not start and nobody wants to dig through ten files to work out why.

That is useful.

It is not judgement. It is not responsibility. It is not experience.

The people still matter most.

Why We Call Him He

This might sound odd from the outside, but we call Theo "he".

That is not because we are pretending he is a human, or because we have forgotten that this is software. It is because inside CreeperHost he is treated as a named assistant with a job, limits, and responsibilities.

He is part of the way the team works now.

But he is not a substitute for the people responsible for him.

That distinction matters to me. It stops Theo from being a faceless automation layer, and it also stops us from pretending that AI decisions somehow float away from human accountability.

If Theo gets something wrong, that is on us to improve.

Not ChatGPT With A CreeperHost Hat

Theo is not trying to be ChatGPT, Claude, or any other large public frontier model.

He is not meant to be the best general chatbot in the world. He is not here to debate every topic, write poetry, or sound like he has read the entire internet.

Theo has a narrower job.

He is built for CreeperPanel, CreeperHost services, and the workflows around running game servers. That means he may sometimes be less polished than a huge public model in open-ended conversation, but that is a trade we are choosing on purpose.

A generic chatbot does not know our panel. It cannot safely inspect service state. It cannot follow our confirmation flows. It should not have broad access to customer systems.

Theo is more limited, but he is useful in the place he is supposed to work.

He can gather facts, explain what he finds, prepare approved panel actions, and stay inside boundaries we control. I would rather have a focused assistant we can understand, operate, restrict, and hold accountable than a more impressive general model running somewhere we do not control.

Local First, Actually

For customer-facing AI features, our position is local-first and CreeperHost-controlled.

That is not marketing language.

I do not want the sensitive part of this running on rented AI hardware in someone else's data centre. I am especially not comfortable with quietly sending customer support tickets, account history, service details, logs, diagnostic output, or chat content off to a third-party hosted LLM provider for processing.

Theo's physical hardware was built by me personally, in-house, for CreeperHost.

He runs in our own data centre, behind our own physical security and two steel doors.

That matters because trust is not only about code. It is about where the data goes, who can touch the machines, who operates them, and who is accountable when something goes wrong.

We may use OpenAI-compatible APIs as a technical interface, because that is a convenient way for software to talk to models. That does not mean customer AI requests are being sent to OpenAI or another hosted AI provider.

For our customer-facing AI features, the work is designed to run on CreeperHost-controlled infrastructure unless we clearly say otherwise.

Safety Before Cleverness

Theo is not given broad destructive access.

He cannot run arbitrary console commands. He cannot delete, edit, rename, unzip, or extract arbitrary files. He cannot destroy services, change root passwords, grant credentials, submit billing changes, cancel accounts, or make serious account decisions.

Where Theo can prepare a service-impacting action, it has to be an allowlisted panel action with a clear confirmation step.

And that confirmation is not just a blind "yes" check.

Service-changing confirmations are reviewed in layers before an action can run, and unclear or mismatched confirmations are blocked. Where model review is needed, two separate local reviewers must both agree that the customer's confirmation matches the requested action. Those reviewers use different models and separate local GPU stacks, including different GPU manufacturers and driver stacks.

If either reviewer disagrees, is unavailable, fails to return a valid decision, or thinks a human should look at it, the action is blocked.

That is not because I think the system is perfect.

It is because I do not think any single part of an AI system should be trusted on its own.

What Happens To Chat History

Ask Theo keeps recent transcript history in your browser's session storage so CreeperPanel can show the conversation if you move around the panel or refresh during the same browser session.

That browser copy is on your device.

On the server side, Theo keeps the active chat session in memory so he can understand the current conversation. If a conversation gets large, older messages and tool records may be compressed into a structured summary, but that summary is still held in RAM as part of the active session.

We do not currently store Ask Theo chat transcripts or compressed session memory in a database for later training, profiling, or review.

If you ask Theo to create something outside the chat, such as a support ticket, diagnostic paste, or attachment, that is different. Those are support or service records because you asked the system to create them.

We do not use customer personal information to train public AI models.

The Waste Part Matters Too

AI should not be wasteful theatre.

Power matters. Cooling matters. Hardware matters. Water matters.

I do not want us burning resources just so we can say we have an AI feature. That would be the worst version of this.

Running our own infrastructure does not give us permission to be careless. It gives us responsibility. We design these systems around bounded context, specific tools, queues, timeouts, and smaller utility models where they make sense.

Useful AI should reduce waste, not create a new kind.

So, What Now?

Theo is new.

He will make mistakes. He will misunderstand things. He will sometimes be less polished than the giant public models people are used to.

We should be honest about that.

But he is ours to build properly.

That means local infrastructure. Clear boundaries. Human responsibility. No quiet shipping of customer AI work to third-party hosted LLMs. No pretending AI is magic. No replacing people who know what they are doing. No careless automation because it looks good in a demo.

We are going to keep exploring this because progress matters.

We are going to keep being careful because people matter more.

• Paul (CEO, Founder, CreeperHost)