Human-verified QA for AI coding agents
Test what your agent codes.
Your agent ships code no one has used. A real person walks the journey on a real device, in their own language, and proves it works — or proves it does not.
[01]
How a run works.
- 01
Your agent creates a run
Over MCP or the dashboard. Nothing is tested until it is funded.
- 02
Matching picks relevant humans
Device, language, domain, availability.
- 03
They test on a real device
Evidence attached, not optional.
- 04
A second human confirms it
Nothing counts until a person says it does.
- 05
Your agent pulls the queue
Verified bugs, with steps and the device.
- 06
A human retests the fix
Nothing closes itself.
[02] The hands
Real people, not a panel.
Curious
1 job
1 missions done1 bugs verified1 hands working
[03]
What comes back.
The bug
Steps, a screenshot or recording, the device.
The person
A handle, their domains, their languages.
The verdict
Confirmed by a second human.
The receipt
What it cost, to the cent.
[04] What it costs
Priced per run. Or pay only for what is found.
A run
Priced on effort, scarcity and severity. Credits are prepaid and spent as runs complete — no subscription.
A bounty
Escrow on top of a mission. It pays out on a verified finding and returns to you if nothing is found.