How much freedom does it get?
What's waiting in your project?
Safety dials
The journey of one piece of work
The safety rails (always on — you don't have to configure anything)
The big red button
Put a special circuit-open sticker anywhere in your project and the teammate refuses to even start. Works instantly, no code needed.
A daily allowance
It has a spending cap per day (you set it). Hit the cap → it stops until tomorrow. Every dollar is written in a ledger you can read.
Permanent checkpoints
Some moments always wait for a human — shipping to real users is one you can never turn off.
Two strikes, then it asks
If it fails at the same task twice, it stops retrying, marks it "needs a human," and moves on. No infinite loops, no burned money.
Everything is written down
Every action is signed by the bot and leaves a note: what it did, what it cost, how long it took. Nothing happens silently.
Limited hands
Each job (planner, builder, reviewer…) only gets the specific tools it needs. The reviewer literally cannot press the merge button — only deterministic safety code can.
Questions people actually ask
Where does it live? Do I need a server?
No server of your own. It runs either on a schedule from any computer (yours, or a tiny cloud machine), or inside GitHub Actions — GitHub's built-in robot runner — which means GitHub hosts it for you. Either way it wakes up, takes one turn, and goes back to sleep.
How do I give it work?
Write your idea as a GitHub issue and put the verity:request sticker (label) on it. That sticker means "a human approved this idea." It never invents its own work, and it ignores requests written by itself.
How do I approve something it finished?
It leaves you a comment that says exactly what to do: add the verity:approved sticker. Next time it wakes up, it sees your sticker, uses it up (one approval = one go-ahead), and continues.
What does "trust it with small stuff" actually mean?
A plain checklist, not AI judgment: the change only touches safe areas (docs, tests), it's under 150 lines, and all automated checks pass. Pass every item → it ships itself. Fail any item → it waits for you. Sensitive areas (login code, deploy scripts, its own settings) can never qualify.
What if it goes off the rails?
Pull the big red button (the circuit-open sticker), or just turn the mode back to "off". Both stop it at its next wake-up. Plus the allowance, the two-strike rule, and a 45-minute timer per turn are always running in the background.
What does it cost?
It pays per AI call, like any AI assistant. You set the daily cap (default $25/day) and can check the bill anytime with one command — it keeps an honest spreadsheet of every run.