Android Viewer

@acertijosapps/android-viewer

Live Android emulator viewer for the browser, plus an MCP server so an AI agent can drive it. Built for the browser pane of AI coding tools (such as Claude Code): you and the agent watch and control the same emulator, in real time, including foldable postures.

• Live video: adb exec-out screenrecord (H.264) over WebSocket, decoded in the browser with WebCodecs.
• Input: click, drag, long press and scroll become adb shell input commands.
• Every AVD on your machine: pick a running emulator or start a stopped one from the viewer.
• Emulator controls: Back, Home, Recents, Power, Volume, Rotate, full-resolution screenshot.
• Foldables: Closed / Half-opened / Opened / Rear display. On emulators they move the simulated hinge, like the emulator's own controls; the stream follows the active display when you fold or rotate.
• MCP server: the same actions as tools, so the agent's taps show up live in the viewer.
• Nothing to install on the device: no APK, no server push; only adb.

Requirements

• Node.js 18 or later.
• Android SDK platform tools (adb) and, to list and start AVDs, the emulator. Found through ANDROID_HOME, ANDROID_SDK_ROOT or the default SDK location; override with ADB and EMULATOR.
• A browser with WebCodecs (Chromium-based browsers, Safari 16.4+, Firefox 130+).
• Works on macOS, Linux and Windows.

Quick start

npx @acertijosapps/android-viewer

Open http://localhost:8093 and pick a device. Deep links: ?avd=Pixel_10 or ?serial=emulator-5554.

android-viewer [options]        Start the viewer

android-viewer mcp [options] Start the MCP server on stdio

-p, --port Viewer port (default 8093, env PORT) --host Interface to bind (default 127.0.0.1, env HOST) --viewer-url Viewer the MCP talks to (default http://localhost:8093)

Other environment variables: MAX_DIM (longest side of the encoded video, default 1024), BIT_RATE (default 6M), SCREENSHOT_MAX (longest side of MCP screenshots, default 1024) and DEBUG=1 (logs every input command).

Use it with an AI agent (MCP)

Claude Code:

claude mcp add android-viewer -- npx -y @acertijosapps/android-viewer mcp

Any other MCP client:

{

"mcpServers": { "android-viewer": { "command": "npx", "args": ["-y", "@acertijosapps/android-viewer", "mcp"] } } }

Also listed in the official MCP Registry as io.github.acertijosapps/android-viewer.

Keep the viewer running (npx @acertijosapps/android-viewer) to watch what the agent does. Without it, the MCP drives the device through adb directly.

Coordinates are normalized to 0–1 over the screenshot (x = pixel_x / image_width), so they hold in any posture or orientation. serial is optional when only one device is connected.

In the Claude Code browser pane

Add the viewer to .claude/launch.json and open it with preview_start android-viewer:

{

"version": "0.0.1", "configurations": [ { "name": "android-viewer", "runtimeExecutable": "npx", "runtimeArgs": ["-y", "@acertijosapps/android-viewer"], "port": 8093, "autoPort": false } ] }

HTTP API

The MCP uses the viewer's API; you can call it too. serial goes in the query string or the body and is optional with a single device. POST requests require content-type: application/json and a local origin, so other websites can't drive your devices.

When a viewer tab is watching the device, actions go through its live session: the viewer updates at once and a watchdog restarts the stream if a gesture produces no frames.

How it works

• One session per device, shared by every tab, stopped when the last one closes.
• Every second it reads dumpsys display (size, rotation and the physical display behind logical display 0) and cmd device_state state. When the display changes, it waits for it to settle and restarts screenrecord with --display-id of the new panel: folding a device switches panels.
• screenrecord only emits frames when the screen changes, so the viewer keeps the GOP since the last IDR for late joiners and sends a screenshot while there is no video.
• Only one screenrecord runs per device at a time: two of them exhaust the emulator's software encoder (err=-38). Restarts requested by clients are rate-limited.
• Screenshots use raw screencap (compressing PNG inside the emulator is slow under load), are scaled down by area averaging and encoded as PNG with node:zlib, with no native dependencies.

Limitations

• Drags are sent on release (one input swipe with the real duration), not live.
• No keyboard passthrough or audio yet; use type_text and press_key.
• Higher latency than a native client such as scrcpy.
• The software encoder slows down with several emulators recording at once; with little free RAM (emulators paged out to swap) every adb shell call can take seconds.
• If an app requests its own posture (DeviceStateManager.requestState), it wins over set_posture while it is in the foreground.
• Flip_Horizontal_Fold style AVDs with a single physical display change state when closed, but the picture stays the same.

Why not ws-scrcpy?

ws-scrcpy was tried first: streaming and input work, but the video freezes on a posture change until you reload (its scrcpy 1.19 server stays bound to the previous display), and it has no posture controls.

Contributing

Branching model (main · beta), Conventional Commits and releases are described in CONTRIBUTING.md; changes are listed in CHANGELOG.md.

Author

Made by AcertijosApps. See the rest of our libraries at acertijosapps.com/librerias.

Questions or proposals: acertijos.contacto@gmail.com or open an issue.

License

MIT