In development
Hangar Box
A desktop console for USB Android farms, where an agent writes the automation once and it then runs without AI.
Real phones, repeatable tests, no model in the loop at run time.
QA on many real Android phones means watching every screen, recording flows by hand and replaying them in parallel. Letting an AI drive every run would be slow, costly and non-deterministic; recording by hand is tedious.
The answer: use the agent to author the automation, then keep it as a normal, deterministic script.
A live grid, a fast screen reader and “New with AI”.
The console shows every phone live, records automations and runs them in parallel with timeouts, retries and schedules, producing JSON/CSV reports with screenshots.
A small server on the phone reads the UI tree directly instead of waiting for uiautomator dump. On top of it, “New with AI”: a Claude agent with tool use observes the screen and acts through the same engine; the session is saved as a regular automation that later runs without AI.
Bounded agent, deterministic replay.
Edits are diffs
“Edit with AI” proposes a diff the person accepts or not.
Step, time and US$ limits
Every AI session is capped and its usage logged.
Secrets stay local
Test-account secrets are never sent to the model; the API key is stored encrypted.
Blocked apps
Social networks and messaging are blocked, also in the normal runner.
Runs without AI
Saved automations replay selector → image → coordinate, with the method in the report.
QA use only
No fraud or evasion features; the terms of use define acceptable use.
Only what has a source.
| What | Value | Source |
|---|---|---|
| Screen read latency | 28 ms median · 40 ms p90 | vs ~2.5 s with uiautomator dump · Redmi 10C · delivery log 2026-10-11 |
| Precise taps on a real phone | 0/5 → 5/5 | Settings app, 5 runs, before/after the scaling fix · delivery log 2026-10-11 |
| Automated tests in the console’s codebase | 90 | Delivery log 2026-10-13 |
| “New with AI” on real devices | Not measured yet | Tested only with a simulated model so far |