I'm Max. For the past year I've been building my own AI assistant — ExadokoAI — in my evenings. It started with a simple annoyance: every AI service I tried hit me with the same wall — "sign up, add a credit card, here's your 3-day trial."
So I built it backwards. You open the site and you can chat right away. No card, no SMS, no phone-number signup.
- Guests get 60 requests and 300,000 tokens per day with no signup at all.
- Signing up with just an email (still free) raises that to 1,500 requests and 10,000,000 tokens per day.
This week the project got its own domain — exadokoai.cc.cd. It's the first real milestone, and I want to share what the thing does and how I got here.
What it does
- Chat. Two model lines: 1.6 (Flash / Basic / Prime) — fast and lively, with a "mood of the day"; 1.7 (Lite / Pro / Pro Max) — heavier tasks, files and images right in the chat.
- Image generation. A studio that draws from a text description — no daily cap.
- Photo understanding. Send a picture and it explains what's in it, reads text from photos, walks through errors on screenshots.
-
Developer API. An OpenAI-compatible endpoint with keys like
exadoko_live_…— plug it into your own clients. - A Telegram bot — up and running.
Limits, no fine print
| Requests / day | Tokens / day | |
|---|---|---|
| Guest (no signup) | 60 | 300,000 |
| After email signup (still free) | 1,500 | 10,000,000 |
Images have no daily limit. Limits reset at 00:00 UTC. Login is an email code with brute-force protection.
Why free?
It's a deliberate bet: first collect real users and honest feedback, then figure out monetization once it's clear what people actually value. Right now the service is cheap to run, and I see no reason to charge for something I can give away. Early-user feedback is worth more to me than revenue.
How I got here
The first version took two evenings. One page, an input, a reply. No history, no settings, ugly layout. But it answered — and the first reply from "your own" service motivates you harder than any tutorial.
Then came personality. If a model answers "like everyone else", nobody will use it. I spent about a month on what's formally called "tone tuning" — really, it's character. That's how the 1.6 and 1.7 lines appeared: the difference has to be audible from the first line, otherwise what's the point of a model picker.
Then came hardening. A free service gets attacked within hours: scripts, code brute-forcing, quota draining. So from day one I built the things you can't demo: brute-force protection, per-minute limits, signed sessions. Boring — but it's exactly what separates a "demo" from a service that doesn't fall over.
Technical bits
- First-token gate. If a model accepts a request and goes silent, we cancel the attempt after ~20 seconds and quietly fail over to the next model. No more staring at a spinner.
- Key rotation across multiple endpoints, with per-endpoint cooldowns.
- Built-in web search. For "what's current" questions — rates, news, prices — the model decides it needs the internet and pulls fresh data itself.
- Stack: Next.js 15 + React 19, Neon Postgres.
Being honest about traction
I won't pretend there's a flood of traffic — there isn't. A handful of users so far. That's normal: almost every indie product starts in a void. The domain is my starting line; now begins the boring part — showing it, fixing it, and telling people about it, again and again.
What's next
Memory across conversations, an agent mode, a mobile interface.
Try it: exadokoai.cc.cd
And if you find a bug — tell me. For a project with barely any users and a lot of enthusiasm, that's the best kind of help.
Top comments (0)