Table of Contents
- Android emulator cloud: choose the service your workflow needs
- Check hardware acceleration before paying for a server
- What the Agent37 Android template includes
- Create a phone and wait for Android readiness
- Give your agent a reliable screen-and-action loop
- Calculate running compute and retained phone storage
- Preserve state deliberately and test the failure paths
- Frequently asked questions
- Is there a free Android emulator cloud?
- Can I use the Play Store on Agent37's default image?
- Can an AI agent control the emulator?
- Where should I start?
Do not index
An Android emulator cloud runs a virtual Android device on remote infrastructure, so your tests or agents can install apps, inspect screens, and send input without tying up a laptop. Choose a managed device service for a ready-made testing workflow, or a VM with hardware virtualization when you need control of the machine. Agent37 provides the second route with an Android template, ADB, persistent phone storage, and browser access.
The buying decision comes down to three questions: does your app run on the supplied Android image, can your automation control it reliably, and what keeps billing between jobs?
This guide covers developer testing and agent execution. Prices and product details were checked on November 17, 2026; the examples are infrastructure estimates, not performance benchmarks.
Android emulator cloud: choose the service your workflow needs
A browser stream, a test farm, and a VM solve different problems. Start with the result you need, then compare products within that category.
Your job | Useful starting point | Check first |
Share an interactive app demo | Appetize | Included minutes and device cap |
Script Android virtual devices | Genymotion SaaS | ADB access and device images |
Run a device test matrix | Google's device testing services | Supported tests and migration |
Keep a phone for your own automation | Agent37 Android VM | App compatibility and sizing |
Appetize offers browser-based app access. Its Free plan lists 30 minutes refilled monthly, one user, and two active devices. That can suit occasional previews before you commit to paid capacity.
Genymotion SaaS supports connecting its virtual devices to ADB through the gmsaas CLI. It fits teams that want managed Android devices while retaining familiar debugging tools.
Google's Firebase Test Lab migration notice now points customers to Developer Device Platform. Test Lab executions remain supported until September 30, 2027. Google's replacement includes Device Run and streaming to physical Android devices; distinguish those physical devices from emulators when choosing a test target.
Agent37 fits a different requirement: a machine your own code controls, with the phone's state retained between runs. You manage your automation and assertions. A VM does not provide a complete device matrix or prove that an app behaves correctly on real hardware.
Check hardware acceleration before paying for a server
An emulator needs more than Linux and a shell. On Linux, hardware acceleration uses KVM, and the process needs permission to use it. CPU acceleration and graphics acceleration are separate requirements: a working KVM device does not promise a dedicated GPU.
The Android acceleration guide documents
emulator -accel-check and warns about running accelerated emulators inside ordinary virtual machines. Do not assume that any VPS or container supports your setup. Choose infrastructure that explicitly supports Android emulation and exposes the required virtualization interface.Agent37's full VM documentation explicitly provides
/dev/kvm for Android emulators. Its standard gVisor sandboxes do not provide that interface. Adding RAM to a standard sandbox will not make it the right runtime for this template.For a custom deployment, check acceleration from inside the actual machine, then boot your selected image and run your app. An advertised CPU count cannot replace that acceptance check. Our microVM and container comparison explains the wider isolation tradeoff.
What the Agent37 Android template includes
The current Android template runs Android 14, API 34, using a
google_apis x86_64 image and a Pixel 6 device profile. It starts the emulator automatically and includes the SDK, ADB, and noVNC browser viewing on port 6901.Google Play services are included. The Play Store app is not. You can sideload an APK with ADB, but you should test any native libraries, login flow, or Google service your app requires against this exact image. A device profile describes an emulated configuration; it does not turn the instance into a physical Pixel.
The template requires a VM runtime. Use
"type": "vm" for the pricing examples below. The documented minimum is 8 GB RAM, because restoring the phone's saved state needs room beyond its configured Android memory. Omitting resources selects 4 vCPU, 8 GB RAM, and 12 GB disk.Requesting more disk gives you space for phone state and artifacts. It does not automatically expand an existing Android data partition. The template sets phone storage when it first creates the AVD, with its documented automatic sizing limits.
This image is a device environment, not an agent harness. It has no chat API and does not bundle the Ubuntu template's Docker daemon. Run your controller elsewhere and call exec, or build a custom image when your workflow needs different software.
Create a phone and wait for Android readiness
Load your workspace key into
AGENT37_API_KEY through your normal secret-management process. With curl installed, this request creates a VM with 4 vCPU, 8 GB RAM, and 20 GB disk:curl --fail-with-body --silent --show-error \
https://api.agent37.com/v1/instances \
-H "Authorization: Bearer ${AGENT37_API_KEY:?Set your workspace key}" \
-H 'Content-Type: application/json' \
-d '{"template":"agent37-android","type":"vm","resources":{"cpu":4,"memory":8,"disk":20}}'Save the returned instance ID as
ANDROID_INSTANCE_ID. Create requires enough wallet balance for a day of the instance's running rate. Eligible new accounts receive $5 starter credit after adding a card; that is one-time credit, not a permanent free tier.A successful create means the machine started. Android may still be booting. The image's
android-wait helper checks Android's boot-completed property:curl --fail-with-body --silent --show-error \
"https://sandbox.agent37.app/v1/instances/${ANDROID_INSTANCE_ID:?Set the returned instance ID}/exec" \
-H "X-Agent37-Key: ${AGENT37_API_KEY:?Set your workspace key}" \
-H 'Content-Type: application/json' \
-d '{"command":"android-wait 600 && adb shell getprop ro.build.version.release"}'Check the JSON response's
exit_code, stderr, and truncated fields. An HTTP 200 can contain a failed shell command. If a long request times out, inspect the device before retrying actions that could run twice. These examples follow the documented API; they are not a claim that we benchmarked a provisioned device for this article.Boot completion is only the first readiness gate. Launch your target app and wait for its expected screen or application state before judging the test. Recheck readiness after every start or recovery.
Give your agent a reliable screen-and-action loop
ADB installs apps, runs device shell commands, and retrieves debugging output. In the Agent37 template it is already connected to the phone, so your backend can drive it through authenticated exec without exposing ADB publicly.
A useful loop captures the current screen, chooses an action, sends that action, and checks the result. The template's
android-shot helper returns a base64 WebP frame, normally 540 pixels wide. Decode stdout only when exit_code is zero and truncated is false. Exec limits each output stream to 512 KB; use SSH for large binary artifacts.Keep screenshot coordinates and device coordinates aligned. A 540-pixel-wide screenshot can be a scaled view of a larger display. Map the position back before using
adb shell input tap, or use element bounds from a UI hierarchy when available. Otherwise, an apparently sensible tap can land on the wrong control.After an action, assert something observable: a package launched, the expected view appeared, or a test produced its result file. Keep a screenshot and relevant logcat output when the assertion fails. Serialize control of each phone so two jobs cannot interfere with the same screen.
For manual inspection, mint a short-lived signed URL for port 6901 and change its path to
/vnc.html, preserving the token. The Android guide documents the exact URL options. Treat the link as a credential because its holder can interact with the phone. The template uses noVNC; do not assume the platform's HTTP edge supports the emulator's UDP-based WebRTC stream.Calculate running compute and retained phone storage
At the current Agent37 rates, VM compute is 2.5 times standard compute. Disk has the same rate. For the 4 vCPU, 8 GB RAM, 20 GB disk example:
Compute per 730 running hours is
(4 × $0.80 + 8 × $0.70) × 2.5 = $22.00. Disk for 730 hours is 20 × $0.09 = $1.80.Usage for one Android VM | Compute | Disk | Total |
Run 60 hours, then delete | $1.81 | $0.15 | $1.96 |
Run 60 hours, keep disk 730 hours | $1.81 | $1.80 | $3.61 |
Run and retain for 730 hours | $22.00 | $1.80 | $23.80 |
Figures round to cents. The retained-disk example assumes ordinary reserved storage, without a cold-storage transition. The default 12 GB disk instead produces $23.08 for 730 running hours. Neither figure is the price of a 2 vCPU standard sandbox, and neither establishes application throughput.
Compute bills while the VM is awake, including idle time. Stop or sleep pauses compute; retained disk continues billing. Auto-sleep is optional. Delete the instance when its saved phone state is no longer needed to end its instance billing. Model calls, external tools, and services your automation uses are separate costs.
For context, Genymotion SaaS pricing lists $0.06 per running device-minute, or $3.60 per hour, on pay-as-you-go. Sixty device-hours is $216 on that plan. Its separate Unlimited offer lists $179 per device per month when billed annually, so heavy users should compare that commitment too. These are different products and resource models, not matched performance measurements.
Small test workloads may favor a free allowance. Firebase Test Lab's published Blaze pricing includes 60 virtual-device test minutes daily, then $1 per device-hour. Its replacement matches Test Lab rates through April 30, 2027, after which pricing changes. Budget against the service and contract you will actually use.
Preserve state deliberately and test the failure paths
Agent37 stores the phone's apps, logins, files, and quick-boot snapshot in its AVD on persistent disk. That data survives the documented stop, start, restart, resize, and image-update operations. Deleting the instance is a different lifecycle event, so export anything you need first.
Persistent files do not guarantee a continuous running session. A VM start or resize cold-boots the machine. Auto-sleep can restore memory, but it can also fall back to a cold boot. Your controller should reconnect, wait for Android, and establish app readiness again.
Check sleep behavior under real automation. Repeated exec calls count as activity, and traffic through a live screen can keep the machine awake. A background job alone does not keep it awake. Choose an idle timeout that fits your work, or disable auto-sleep while a job requires uninterrupted execution.
Before scaling, install your actual APK, exercise its critical flow, restart the VM, and verify recovery. Measure completion time and resource pressure with that workload. Test real hardware separately when the result depends on physical sensors, vendor behavior, or device security properties.
Frequently asked questions
Is there a free Android emulator cloud?
Some services include limited free minutes. Appetize lists 30 minutes monthly; Test Lab has testing allowances subject to its plan and migration. Agent37 offers eligible accounts one-time starter credit after adding a card. Ongoing VM and retained-disk usage is metered.
Can I use the Play Store on Agent37's default image?
The default image includes Google Play services, not the Play Store. Use APK sideloading, or follow the Android documentation to build a suitable custom image and matching AVD. Installing another SDK package alone does not replace an existing phone.
Can an AI agent control the emulator?
Yes. Your controller can call exec to capture screenshots and send ADB input. You supply the reasoning model, control loop, and success checks; the Android template does not provide a chat API.
Where should I start?
Choose one representative app flow and one phone. Create an Android VM in Agent37, follow the Android setup guide, and verify boot, app readiness, screenshots, and recovery. Use that measured run to choose your final shape and running schedule.
