Can DeviceyAI do anything I could not do myself?

No. It runs in your browser tab over the same ADB connection your session already holds, so its reach is exactly your reach. The phones are bootloader-locked with verified boot enforcing and no su binary, so neither you nor DeviceyAI has root.

Does it touch the phone without asking?

Not unless you let it. Your first session starts read-only: it can look at the screen and nothing else. Tapping, typing, launching apps and installing your APK need the 'Let it act' toggle, and after you have turned it on once we open your later sessions the same way rather than asking again. Shell access is a second, separate toggle on top of that, and it is never remembered — you turn it on per session.

DeviceyAI bills in AI credits, separate from device time. One credit is €0.01. A small task — open Settings and read back the Android version — cost 3 to 16 credits on real hardware depending on the model, so a few cents. Paid plans include a monthly allowance and you can top up in packs.

Can I use my own API key?

Yes. Add a provider key in Account → AI and your DeviceyAI turns bill to your own provider account instead of spending credits. We still record the usage so you can see what ran.

Can it run a task while I am away?

No, and that is deliberate. DeviceyAI only works inside a session you have open, so a person is always watching the mirror while a physical phone is being driven. There is no scheduler and no unattended runner.

Which models can I choose?

Today: Claude Opus 5 and Claude Sonnet 5. Every model id is verified against the provider's live model list before it is offered, so a model in the dropdown is one that actually answers.

DeviceRent

Live for paid plans

Tell it what to do.
Watch it do it on a real phone.

DeviceyAI is an agent that drives the Android phone you rented. It reads the screen, taps, types, and installs the APK you uploaded — through the same connection you use by hand, with every step visible in the live mirror.

Start a session See pricing

tool calls to open Settings and read back the Android version, on a real Pixel 7a: 12
tool calls to open Settings and read back the Android version, on a real Pixel 7a
AI credits that run cost — about 16 cents: 16
AI credits that run cost — about 16 cents
tools it can use, all of them things you can do by hand: 24
tools it can use, all of them things you can do by hand
new ways into the rack — it adds no network path: 0
new ways into the rack — it adds no network path

A run

You describe the task. It works the phone.

Each step is a real action on real hardware: read the screen, decide, tap, read again to check what happened. You see the tool names as they run, the screenshots it took, and what it concluded.

Open Settings and tell me the Android version
observe_screen12 interactable of 48 nodes — home screen
launch_appcom.android.settings — started
observe_screenNetwork · Connected devices · Apps · … · About phone
tapx=540 y=1893
observe_screenDevice name · Legal · Android version · Model
tapx=540 y=1104
observe_screenAndroid version 16 · Security update · Build number
This Pixel 7a is on Android 16. I opened Settings, scrolled to About phone, and read it off the Android version row.

condensed from the acceptance run · 12 tool calls · 16 credits

Capabilities

Three levels, and you hold both switches.

A fresh session can look and nothing more. You decide whether it may touch the phone, and separately whether it may use the shell. Turning actions off turns the shell off with it.

Look

always on

observe_screen
Read the screen as a list of elements with exact positions
screenshot
Take a picture of the screen and look at it
list_packages
List the apps installed on the phone
device_info
Read the model, Android version, screen size and orientation
wait
Pause while something loads or animates
list_instrumentations
Find the exact test-runner target installed on the phone
logcat_tail
Read the tail of the device log, filtered to what matters
read_device_file
Read a text file a test wrote — a JUnit XML or coverage report

Act

your toggle

tap
Tap a point on the screen
swipe
Swipe or scroll
type_text
Type into the focused field
press_key
Press Home, Back, Recents, Power, Volume, Enter or Delete
launch_app
Open an installed app
install_apk
Install an APK you uploaded to this rental
list_uploads
See which of your files are available to install
run_instrumentation
Run an instrumentation suite already on the phone and report pass/fail
save_test_file
Keep a test file it wrote, for you to download

Shell

a second toggle

shell
Run an ADB shell command

Appium

with actions on

appium_start_session
Open a real Appium (UiAutomator2) session on the phone
appium_page_source
Read the UI hierarchy through that session
appium_find
Prove a locator resolves on the live app before it goes into a test
appium_click
Click an element it found, rather than a raw coordinate
appium_send_keys
Type into an element it found
appium_end_session
Close the session and release the phone

How it works

The model never touches the phone.

It runs on our server and can only ask for tools. The tools run in your browser tab, over the ADB connection that tab already holds for your rented device. Nothing new listens, and nothing new reaches the rack.

The model never reaches the phone, and neither does our server. Tools run in the tab you already have open, over the connection it already holds — so DeviceyAI can do exactly what you can do by hand, and nothing else.

Limits

What it cannot do.

Being straight about the edges is more useful than a longer feature list.

No root

Every phone in the rack is bootloader-locked with verified boot enforcing and no su binary. There is no Magisk, no Xposed, and no system-level modification — for you or for DeviceyAI.

No unattended runs

DeviceyAI only works inside a session you have open. There is no scheduler and no background runner, so a person is always watching while a physical phone is driven.

Nothing that bricks a phone

Reboots to bootloader or recovery, factory resets, disabling ADB and removing the packages the rack depends on are refused before they reach the device.

It stops on its own

One instruction is capped at 40 steps and a session at 500 credits, so a confused run ends and asks rather than spending your balance. Stop cancels mid-generation.

Credits

Priced per turn, not per seat.

AI credits are separate from device time and are charged from what a turn actually used — never a flat fee, and nothing at all if the provider call fails. One credit is €0.01.

Starter

€5

500 credits

Standard

€12

1,500 credits

Bulk

€30

4,000 credits

Every paid plan includes a monthly allowance

300
mid
1,000
pro
2,000
dedicated slot
2,000
enterprise

Prefer to bill your own provider account? Add an API key in Account → AI and DeviceyAI spends no credits at all.

Models

Only models we have actually run.

Every id in the dropdown has been checked against the provider's live model list and driven through a real task on real hardware. A model that looks right on paper but fails on the second turn does not get listed.

Claude Opus 5

1,000,000 token context

Claude Sonnet 5

1,000,000 token context

Questions

The things people ask first.

Can DeviceyAI do anything I could not do myself?: No. It runs in your browser tab over the same ADB connection your session already holds, so its reach is exactly your reach. The phones are bootloader-locked with verified boot enforcing and no su binary, so neither you nor DeviceyAI has root.
Does it touch the phone without asking?: Not unless you let it. Your first session starts read-only: it can look at the screen and nothing else. Tapping, typing, launching apps and installing your APK need the 'Let it act' toggle, and after you have turned it on once we open your later sessions the same way rather than asking again. Shell access is a second, separate toggle on top of that, and it is never remembered — you turn it on per session.
What does it cost?: DeviceyAI bills in AI credits, separate from device time. One credit is €0.01. A small task — open Settings and read back the Android version — cost 3 to 16 credits on real hardware depending on the model, so a few cents. Paid plans include a monthly allowance and you can top up in packs.
Can I use my own API key?: Yes. Add a provider key in Account → AI and your DeviceyAI turns bill to your own provider account instead of spending credits. We still record the usage so you can see what ran.
Can it run a task while I am away?: No, and that is deliberate. DeviceyAI only works inside a session you have open, so a person is always watching the mirror while a physical phone is being driven. There is no scheduler and no unattended runner.
Which models can I choose?: Today: Claude Opus 5 and Claude Sonnet 5. Every model id is verified against the provider's live model list before it is offered, so a model in the dropdown is one that actually answers.

Rent a phone. Then stop driving it yourself.

DeviceyAI is included on every paid plan with a monthly credit allowance. Bring your own API key and it costs nothing extra at all.

Create an account How the session works

Early access

The fleet is small, and you decide what grows it.

Six real Pixel 7a phones, hosted in the Netherlands — that is the whole rack, and the live count above is not marketing. Tell us what you need a real device for: we'll email you when one matching it is free, and your answer decides which phones get bought next.

Want the longer version? See what we can and can't do yet.