limrun-android-emulator

Limrun Android Emulator

Interact with an app running on a Limrun cloud Android emulator, from any environment (Linux, Windows, macOS, VM, container). This skill is build-agnostic: it assumes the APK was already built, usually by limrun-gradle. Keep build concerns in that skill; this one is about driving the running emulator.

Never use a local emulator, a local Android SDK, or Android Studio.

Auth and CLI

Install if needed: npm install --global lim. Auth is lim login or LIM_API_KEY (it may be set outside the project, so don't ask for it just because it's missing from .env or the shell). The CLI is the source of truth: the commands in this skill are verified, but if a flag errors or you need one not shown here, check lim android <subcommand> --help instead of guessing.

Installing an app

You can build with Limrun's remote Gradle service and have the APK installed on a fresh emulator automatically, or install a pre-built local APK or a URL. One default to know first: lim android create opens an ADB tunnel (--connect) and a browser tab with the live stream (--open) unless told otherwise. As an agent, pass --no-open always, and --no-connect unless you need adb right away.

Build and install

Upload the APK built by limrun-gradle as a named asset, then create an emulator with it pre-installed:

lim gradle build . --upload myapp.apk
lim android create --install-asset myapp.apk --no-open --no-connect

Create blocks until the instance is ready to drive; no boot wait is needed. The create output includes a signed stream URL; share it with the user as a Markdown link, like [Live emulator](<signed-stream-url>). If you have a browser the user can see, open the URL there and tell them. Create also prints a console URL: it opens the same live view but requires a console login, so prefer the signed stream URL for sharing.

Useful create flags: --reuse-if-exists (reuse a running instance with the same labels), --rm (delete when the CLI exits), --jurisdiction us|eu|as (where the instance runs; don't use --region, it is deprecated), --inactivity-timeout / --hard-timeout, --display-name, and --label k=v (find labeled instances later with lim android list --label-selector k=v).

Install app from local

Create a new emulator, then install a local file or a URL:

A local path is uploaded to Limrun Asset Storage first; a URL is fetched by the instance itself, which is much faster for large APKs than uploading from your machine. install-app returns as soon as the app is sent; the install finishes in the background within seconds. Newly installed apps land in the app drawer, not the home screen, so don't look for their icon; just launch the app with lim android launch-app <package> --detach (without --detach it blocks watching the app until it exits), or confirm with lim android adb-shell -- sh -c "pm list packages | grep <name>".

Every time you need to install a new version of the APK, sync it instead of reinstalling:

It sends only a delta against the APK already on the instance, then reinstalls. --watch keeps re-syncing on file changes, and --launch-mode ForegroundIfRunning|RelaunchIfRunning controls what happens to the running app after each install.

Targeting the right instance

Most lim android commands default to the last created instance and resolve the "current" one from the git repo / worktree of your cwd. In a different directory (or outside any git repo) a command can report that no recent instance was found even though one is running. The reliable recipe:

Once you have the ID (format android_<region>_<ulid>), pass --id <android-instance-id> to all lim android calls for the rest of the session. Alternatively, git init the project so the workspace resolves on its own. When controlling multiple instances, always pass --id.

Launching the app

Launch and stop installed apps by package name:

Without --detach, launch-app blocks watching the app: when it crashes, ANRs, or is stopped, the command prints the exit reason, crash details with the stack trace, and a recent app log tail, then returns. That report is the way to see why an app died without adb; for logs while the app is running, use lim android app-log (below). There is no list-apps; discover package names with lim android adb-shell -- pm list packages, or take the application ID from the build.

App logs

One app's logs need no tunnel:

Shell and files

One-shot shell commands and file transfer need no tunnel either:

They run with the same permissions the adb shell user has, and adb-shell exits with the command's exit code.

Full logcat and interactive adb over the tunnel

Full-device logcat and anything interactive (Android Studio, scrcpy, streaming) go through plain adb over the CLI's tunnel. Start the tunnel in a background shell and keep it alive:

connect runs adb connect for you (use --adb-path if adb isn't on PATH). The printed 127.0.0.1:<port> is the device serial; pass it with -s to every adb call (adb devices also lists it):

The tunnel lives and dies with the process that started it: when that shell exits, adb devices shows the serial as offline while the instance keeps running. Just run lim android connect again to get a new tunnel (the port changes each time). Stale offline serials are harmless; adb disconnect clears them.

Testing changes

When emulator interaction is part of the task, test new or changed functionality with the interaction commands after each install or sync. Focus on what changed, plus a quick smoke test of core flows. Start by reading the element tree to see what's on screen before acting:

The output is the raw UIAutomator XML hierarchy on a single line, so a plain grep echoes the whole document. Split it into one node per line first, then grep for the text, resource-id, content-desc, or bounds you need rather than dumping the whole tree into context:

Interacting with the app

Prefer tapping by resource id, then by visible text or content description, then coordinates as a last resort:

Selector values match exactly, not by substring: --text "Save" does not match a "Save draft" button. Copy the value verbatim from element-tree. A selector that matches nothing fails in a couple of seconds with No element found for selector. Other selectors: --class-name, --package-name, --index, --clickable, --enabled, --focused, and --bounds-contains-x/y; combine them to narrow a match. On web pages in the browser, resource ids are the page's own DOM ids (like searchIcon) and are often empty; select by --text plus --class-name there, or fall back to coordinates from the node's bounds. To inspect matches without tapping, use find-element with the same selectors:

For text input, target the field directly; no prior focus tap is needed. type takes the same selectors as tap-element (--class-name android.widget.EditText --focused works for a field with no resource id):

For scrolling and navigation:

After every interaction, re-run element-tree to confirm the UI transitioned. No sleep is needed between a tap and element-tree; the tap blocks until done. A page load after open-url is asynchronous though: re-run element-tree until the node you expect appears. A present but childless android.webkit.WebView means the page is still loading, not a broken tree.

When the element tree is empty

Some React Native and Expo apps expose no accessibility nodes at all, which leaves element-tree, tap-element, and find-element blind (system dialogs still expose nodes). Fall back to driving by pixels: take a screenshot, read the coordinates of the target, and use tap x y / type --x --y. Screenshot pixels map 1:1 to tap coordinates, so a button centered at (360, 1322) in the image is tapped with lim android tap 360 1322. Re-screenshot after each action to confirm the result.

Screenshots and video

Screenshot takes a positional path (not -o):

Use the element tree for functional assertions (element existence, text, state changes) and screenshots only for visual properties. For anything involving motion (animations, gameplay, streaming UI), prefer video:

record stop accepts --quality 5-10. Recorded frames are half the screenshot resolution, so read tap coordinates from screenshots, never from video frames. For UI changes, include a demo video in the pull request so the user can see it.

Simulate the microphone with an audio file

For voice-driven flows (assistants, speech-to-text, audio calls), play a local audio file as the emulator's microphone. The app hears the audio through its normal capture pipeline:

WAV and MP3 work. The file is pushed over adb, so this command needs a local adb binary (--adb-path if it's not on PATH) and opens its own short-lived tunnel. Camera injection is iOS-only; for camera-driven test flows use limrun-ios-simulator.

Shape network bandwidth

Test slow-network behavior by capping the instance's Wi-Fi bandwidth:

Tunnel the app's traffic through your machine

When the app must reach a service only your machine can reach (a local dev server, a VPN-only staging API), or you need to see its HTTP traffic, start a destination tunnel. Only the destinations you select are rerouted through the machine running lim; everything else leaves the instance directly.

• An exact selector (localhost:port or IP:port, port >= 1024) becomes a listener on the emulator, also reachable as 10.0.2.2:<port>; the app's connections to it land on your machine and are dialed there.
• Domain selectors (api.example.com, "*.corp.example") are intercepted on the emulator and dialed from your machine, so your DNS and VPN apply. Apps that resolve DNS themselves over HTTPS bypass domain interception.
• Start the tunnel before launching the app: connections opened earlier keep their original route. One tunnel per instance; a second start fails.

As an agent, always pass --detach: it returns once the tunnel is READY and keeps it alive in a background process. Manage it with:

Inspect HTTP traffic, capture HAR, persist a network log

Inspection is on by default: every HTTP and HTTPS request through the tunnel is decoded, printed as one summary line per request (in the tunnel log file when detached), and shown live in the console's network panel.

--persist uploads a body-inclusive network log as a session artifact when the tunnel stops or the instance terminates; it appears on the instance's session page in the console with a HAR download (default lifetime 3 days, --ttl <seconds> up to 30 days). HTTPS is decoded with an emulator-trusted CA, so apps with certificate pinning fail through inspected domain selectors: leave the pinned host out of the selectors or pass --no-inspect to relay bytes opaquely (no summaries, HAR, or persistence).

Preview URL for humans

Upload the APK to Limrun Asset Storage and return a preview URL for the user to open and test the app manually in the browser:

Opening the link in the Limrun console provisions an emulator with the APK pre-installed.

Cleanup

When the work is completed, you can delete the emulator. delete takes a positional ID (--id is not a valid flag here, unlike other commands):

Gotchas

• The fleet is x86_64. Emulators report x86_64,arm64-v8a ABIs and run arm64 code through translation, but an APK whose native libraries are arm64-only for some vendor SDKs installs fine and then crashes with UnsatisfiedLinkError when that code first loads. Build with x86_64 native libs included.
• Selectors match exactly. tap-element --text and find-element --text need the full, exact string from element-tree; substrings match nothing.
• install-app returns before the install finishes. The app lands a few seconds later; verify with find-element or pm list packages before launching. Prefer install-app <URL> or create --install-asset over raw adb install: a big APK over adb install streams with zero progress output and looks hung for minutes, and killing it mid-stream corrupts the install.
• The ADB tunnel is session-bound. It dies with the shell that started it while the instance keeps running; reconnect with lim android connect and re-read the port, it changes every time.
• Failed create pipes can still leak an instance. If a create invocation errors client-side (broken pipe, JSON parse), check lim android list; the instance may exist anyway and should be deleted.
• Empty element tree usually means a React Native app, not a broken instance. See "When the element tree is empty" above.
• element-tree can be large. Pipe through grep to extract what you need rather than dumping the whole tree into context.
• Instance resolution can miss in a non-git dir. See "Targeting the right instance" above; pass --id when in doubt.
• Build errors are the build skill's job. If the APK isn't building, the failure is upstream; go back to limrun-gradle.