Free AI SystemYour own uncensored AI,
on your own domain
An abliterated model has had its refusal reflex surgically removed. This free system puts one on your own server, behind your own login, for about €7 a month. Claude Code does the whole build while you watch.
You already know the feeling
You ask a flagship model something perfectly reasonable and it decides not to help. Not because the answer is dangerous — because the topic tripped a guardrail.
you
Walk me through how this exploit actually works.
assistant
I can't help with that.
Security research. Fiction with adult themes. Blunt competitive and negotiation analysis. Medical, legal and financial questions where the hedging is noise. Whole categories of ordinary work get declined out of caution rather than because there is actually a problem — and there is no appeal, no context you can add, no way through.
Run a model that doesn't have the reflex
Open-weight models ship with a learned refusal behaviour — a specific direction in the model's internal state that fires when it decides to decline. Abliteration is weight surgery that removes that direction. Nothing is retrained and the model's knowledge is unchanged. It simply loses the reflex to say no.
- A ChatGPT-style chat interface at chat.yourdomain.com
- A login on the front door, so only your email address gets in
- No inbound ports open at all — nothing on the box to attack
- No per-message cost. Same price at five messages a month or five thousand
- An API endpoint you can call from your own scripts, n8n, or Claude Code
You don't need to know Docker, Linux, or any of the APIs involved. You need two accounts and about twenty minutes. Claude Code asks for what it needs, one thing at a time, and stops before spending a cent if something is missing.
Eight files. One install.
Drop the folder into your Claude Code skills directory, restart, and say “set up my own uncensored AI server.”
README.md
Install it, what it costs, and the honest version of what you're getting.
SKILL.md
The seven-phase workflow Claude follows, start to finish.
hardware-and-models.md
Which box to buy, which model fits it, real speed expectations.
provisioning.md
Firewall first, server second — so the box is never naked.
stack-install.md
The Docker stack, and why every setting is what it is.
cloudflare-tunnel.md
Tunnel, DNS, and the login gate on your own domain.
verification.md
The seven checks that prove it actually works.
gotchas.md
Every failure that looks like success. The most valuable file here.
gotchas.md is the part that took real deployments to learn. Empty responses that look like a broken model. A model picker that stays empty while inference is perfectly healthy. Settings that stop responding to your config file after the first boot. Every one of them wastes an afternoon if you don't know it's coming.
What this is not
This runs on a CPU, not a GPU
Roughly 8–12 words per second. Comfortable to read as it types, noticeably slower than ChatGPT. No cloud host in this price range sells a GPU.
Text only
No image or video generation — those need a GPU, which is a different machine at roughly 25× the price. The system tells you your options if you want that later.
It changes the model, not the law
Removing a refusal doesn't make anything legal that wasn't. It also removes hedging, so these models are more willing to be confidently wrong. Treat output as a fast draft, not an authority.
Everything in the pack is built around that reality rather than glossing over it. It's your server and your responsibility.
Take it. It's free.
One download, one Claude Code session, and you have a private model that answers the question you actually asked.
Want the rest of what we run — every AI system, skill and build? AI Automation Insiders is where it lives.