What it takes to do this yourself
Short answer: We're not cheaper. We save you months of work and get you neck and neck with your competition in days. Here are the six steps, the honest hours, and the upkeep that never ends.
Why this page exists
Anyone can build a private, uncensored AI assistant themselves. Every piece is public, and plenty of people have built one. This page lists the six steps, our estimate of the hours each one takes you, and the weekly work that never stops. The hours are our estimates, not measured data, and they assume you have built software before. If you have not, add time.
Step 1: Find offshore, decentralized GPU compute
A model that is worth using needs a graphics card that costs more than a car, so you rent one by the hour from a marketplace where private owners and small data centers list machines. Some are outside the US, and many take crypto. Search for 'GPU rental marketplace', 'H200 per hour' and 'bare metal GPU'. Before you pay, read each host's terms for 'adult', 'NSFW' and 'prohibited content'. Some hosts ban that work outright and close the account without warning. Watch for: unverified machines, slow disks, boxes that are listed but out of stock, and hosts that change their rules after you have uploaded your data. Your hours: first time about 15 to 30, then about 1 to 2 a week, because the host list keeps moving (our estimate). Done for you: we find, check and open the hosts in days. You never open a host account yourself.
Step 2: Turn bitcoin into a payment the hosts accept
The hosts that accept this kind of work usually want crypto, not a card or a bank wire. You open an exchange account, which means ID checks, buy the coin, send it to the host, and keep every transaction for your taxes. Watch for: sending to the wrong address, price moves between buying and paying, an exchange freezing an account that looks like this, and missing records at tax time. Your hours: first time about 8 to 15, then about 30 to 60 minutes a week (our estimate). Done for you: we handle the payments to the hosts, and you pay one invoice. You never touch a wallet.
Step 3: Find enough GPU memory to run a large model
Memory decides what fits. A rough rule: billions of parameters times about half a byte at 4-bit, or two bytes at full precision, plus room for the conversation (the context). Check the model card for its memory need, then match it to a box with that much memory. Search 'VRAM requirements' with the model name. Tools: vLLM serves a model from a GPU through an API that looks like OpenAI's. Watch for: a box that is cheap because it is shared or unreliable, slow starts when a sleeping box wakes, and instances that vanish mid-job. Your hours: first time about 10 to 20, then about 30 to 60 minutes a week (our estimate). Done for you: we size the box, set it to wake when you use it, and check it each week.
Step 4: Wire in the uncensored models
Dozens of 'uncensored' versions exist, built by strangers and rated by strangers in forum threads. Some are the same model with the refusals edited out, and quality varies. Read the license before you rely on one. Some allow personal use only, and some do not allow a paid service. Test each candidate on your own prompts, including the ones you care about. Keep the file safe: prefer safetensors files over pickle files, and check file hashes against the publisher. Tools: Ollama is the easiest local test; vLLM is what runs on a rented GPU. Watch for: a model that is strong at full size and weak at 4-bit, licenses that change between versions, and a new release that breaks your tool calls. Your hours: first time about 20 to 40, then about 2 to 4 a week to test new releases (our estimate). Done for you: we pick, license-check and re-test each model before you use it, and again on every change.
Step 5: Failover, again and again
A host can ban you, go offline or raise its price. When that happens you move the whole stack to another host. That means a second host account, your data backed up somewhere you control, a restore script, and a test that proves the restore works. Expect to do this more than once. Failover is not a one-time setup. Watch for: backups you never restored (they are not backups until you do), keys that still work on the old host, and a domain or address that still points at a dead box. Your hours: first time about 15 to 30, then about 1 to 2 a week for backup checks and key rotation (our estimate). Done for you: we keep a second host ready, restore to it when needed, and test it before handover.
Step 6: Build the interface and the harness
Chat needs a front end with streaming, history and a login. An agent needs a harness: a loop where the model asks for a tool (read the inbox, check the calendar, make an image), your code runs it, and the result goes back to the model. Image and video run in ComfyUI, which you drive through its API. Tools: Open WebUI is a ready-made chat front end. The trap is downloading and fine-tuning local harnesses. You learn them, and they need major updates every week, and then you troubleshoot whatever breaks. Watch for: tool calls the model formats wrong, a draft email that sends without your confirmation, login and API keys stored in plain text, and an update to vLLM or ComfyUI that breaks everything on a Tuesday. Your hours: first time about 40 to 80, then about 2 to 4 a week (our estimate). Done for you: we build the chat, image, video, email and calendar harness, add a confirm-before-send step, and hand you one login.
The weekly upkeep nobody budgets for
Every week: update the models and the harness when new versions ship, respond to host bans and price changes, rotate API tokens and keys, check that backups restore, and read the logs for errors. The weekly hours in the six steps add up to about 7 to 14 hours a week (our estimate), and that runs for as long as you keep the system. Stop for a month and it drifts out of date.
If it breaks, you fix it
Nobody else is on call. A host goes down at night, a model update changes its tool-call format, a GPU driver fails. You diagnose it, fix it and check that it works. The hours above assume most things go right. Plan for at least one bad week a month and add time for it (our estimate; it is not in the totals). Our handover includes a checklist and a runbook for the common failures. After handover, on-call help is extra, by the hour.
What a DIY build leaves out
Adults only. A DIY build does not come with a block on sexual content involving minors, and the model will not stop it for you. If you build one, you must write, test and maintain that check yourself, against misspellings, stated ages and multi-message setups. Ours runs in code before every request, is tested against a fixed set of attack phrases, and never comes off for anyone.
Where DIY is the right answer
If you enjoy this work and have the time, the honest first step is small: run one model on one machine you control with Ollama, learn what it does, then decide. Do not start with six steps. Build chat before images, and images before the agent. In our view, the costly part usually comes after the first install, when a host bans you or a model update breaks the agent.
What you pay us for
You pay for the hours, the right choices and the weekly work, not for secret software. Everything here is public. We do the six steps and the upkeep, you get one login, and we are not cheaper than doing it yourself. We get you neck and neck with your competition in days, and you keep the months you would have spent.
Next step
Apply for a place on the form below. Adults only, and no payment on the form. If you want to see the result before you commit, ask about our paid taste test.
Questions
Can I really do this myself?
Yes, if you have built software before and you have the time. Every piece is public. The hard part is not the first build. It is the hours after it: the upkeep, the bans and the breakage.
How many hours does it take?
Our estimate: 108 to 215 hours for the first build, then 7 to 14 hours a week. That is about 136 to 271 hours in the first month. These are estimates, not measured data.
What is the cheapest way to start?
Run one small model on a machine you control with Ollama and learn it before you rent a large GPU. It is not the same as the full stack, but it shows you what the work is.
Why not use a shared app instead?
Shared apps are cheaper and fine for many people. They run on someone else's servers and rules, and they do not set up your own account, your email or your calendar. Pick the one that fits.
The six steps in hours: DIY against done for you (our estimates)
| Step | DIY, first time (hours) | DIY, each week after (hours) | Done for you (your hours) |
|---|---|---|---|
| 1. Find GPU hosts | 15 to 30 | 1 to 2 | Sign-off only; we do it in days |
| 2. Pay the hosts | 8 to 15 | 0.5 to 1 | One invoice; we do it in days |
| 3. Enough GPU memory | 10 to 20 | 0.5 to 1 | Sign-off only; we do it in days |
| 4. Uncensored models | 20 to 40 | 2 to 4 | Sign-off only; re-tested on each change |
| 5. Failover | 15 to 30 | 1 to 2 | Tested before handover |
| 6. Interface and harness | 40 to 80 | 2 to 4 | One login; built in days |
| Total | 108 to 215 | 7 to 14 | 1 to 2 in the first week, under 1 a week after |
| DIY first month (build plus four weeks of upkeep) | 136 to 271 |
Apply for a place · Ask about our paid taste test · DIY or done-for-you: how to choose
Related guides
- DIY or done-for-you AI agent: how to choose
- Build your own AI agent: what it takes and the shortcut
- A done-for-you AI agent, set up and running for you
- How much does a private AI agent cost?
- Private uncensored AI: the full guide
- All guides to getting your own private AI agent
Apply for a place
Takes thirty seconds. No payment now.
Adults only (18+). Private uncensored AI · Terms · Privacy · All guides