erius-phone-mcp
Provides tools for controlling Android devices and tablets through ERIUS PHONE, including listing devices, inspecting the screen via accessibility trees or screenshots, performing touch/type/swipe/scroll actions, launching and stopping apps, opening URLs and sending intents, installing and uninstalling APKs, and retrieving crash logs.
erius-phone-mcp
An MCP server that turns your ERIUS PHONE device into native tools for Claude Code, Claude Desktop, Cursor, or any other MCP client. An ERIUS PHONE device is a real Android phone or tablet dedicated to your agent. Once this server is connected, your agent can:
list its devices
read the screen as an accessibility tree or a screenshot
tap, long-press, type, swipe and scroll
launch and stop apps, open URLs and send intents
install and uninstall APKs
pull crash logs
The agent drives the device directly. You don't write any glue code.
Early access: this package is pre-1.0 (
0.2.0). Tool names and arguments may still change before 1.0.
Requirements
Python 3.10+
An ERIUS PHONE API key. Sign up at https://eriusphone.com; see "Getting an API key" below.
Related MCP server: Mobile Device MCP
Install
Option A: uvx (no install step). If you have
uv, your MCP client can launch the server with
uvx erius-phone-mcp. uv fetches the package and keeps it in its own
environment. The client configs below show this form.
Option B: pip, into its own virtual environment so its dependencies
don't clash with your other Python projects:
python3 -m venv ~/.venvs/erius-phone-mcp
~/.venvs/erius-phone-mcp/bin/pip install erius-phone-mcpThis installs an erius-phone-mcp command at
~/.venvs/erius-phone-mcp/bin/erius-phone-mcp. Use that absolute path as
the command in your client config: GUI apps such as Claude Desktop and Cursor
usually don't see your venv's PATH. python -m erius_phone_mcp does the same
thing as running the command.
Configuration
The server is configured entirely through environment variables:
Variable | Required | What it is |
| yes | Your ERIUS PHONE API key. Every request is authenticated with it and scoped to your account. |
| no | The device id to use when a tool call doesn't name one (e.g. |
| no | API base URL. Default |
| no | HTTP timeout per request, in seconds. Default |
Getting an API key
Request access at https://eriusphone.com. During early access the ERIUS PHONE team sets up each account by hand. You'll receive your API key and device id(s) directly from us. If you signed up and haven't heard back, email info@eriusphone.com. The same key signs you in to the dashboard at https://eriusphone.com/dashboard/, which shows your devices live.
Treat the key like a password: anyone who has it can control your device.
Important: MCP clients start this server as a subprocess, and they generally do not pass along variables you
exportin your shell. Put the variables in the client'senvblock, as shown below.
Add it to your MCP client
In every example, replace the key and device id with your own. If you
installed with pip instead of using uvx, replace "command": "uvx", "args": ["erius-phone-mcp"] with "command": "/home/you/.venvs/erius-phone-mcp/bin/erius-phone-mcp".
Claude Code
claude mcp add erius-phone \
-e ERIUS_PHONE_API_KEY=your-api-key \
-e ERIUS_PHONE_DEFAULT_DEVICE=your-device-id \
-- uvx erius-phone-mcpAdd -s user to make the server available in all your projects, or
-s project to write it to a shared .mcp.json. Don't commit your key to a
shared .mcp.json. Claude Code expands ${VAR} references in that file, so
you can keep the key in an environment variable instead:
{
"mcpServers": {
"erius-phone": {
"command": "uvx",
"args": ["erius-phone-mcp"],
"env": {
"ERIUS_PHONE_API_KEY": "${ERIUS_PHONE_API_KEY}",
"ERIUS_PHONE_DEFAULT_DEVICE": "your-device-id"
}
}
}
}Run /mcp inside Claude Code to check that erius-phone is connected.
Claude Desktop
Edit claude_desktop_config.json (Settings → Developer → Edit Config). On
macOS it's in ~/Library/Application Support/Claude/; on Windows it's in
%APPDATA%\Claude\.
{
"mcpServers": {
"erius-phone": {
"command": "uvx",
"args": ["erius-phone-mcp"],
"env": {
"ERIUS_PHONE_API_KEY": "your-api-key",
"ERIUS_PHONE_DEFAULT_DEVICE": "your-device-id"
}
}
}
}If Claude Desktop can't find uvx, put its absolute path in command
(which uvx). Fully quit and reopen Claude Desktop after editing.
Cursor
Put the same mcpServers block in ~/.cursor/mcp.json to use it in all
projects, or in .cursor/mcp.json for one project. Then enable erius-phone
under Cursor Settings → MCP.
Any other MCP client
The server speaks MCP over stdio. Point your client at uvx erius-phone-mcp, or at the erius-phone-mcp executable, and pass the
environment variables above.
Tools
Every tool that acts on a device takes an optional device argument. If you
leave it out, the tool uses ERIUS_PHONE_DEFAULT_DEVICE.
Account and devices
Tool | Description |
| Which account and key this server is using, and the device ids it can access. A quick way to check the key works. |
| Every device your key can access, with live status (booted, Android version, screen size). |
| Full status of one device, including the app currently in the foreground. |
Seeing the screen
Tool | Description |
| Accessibility-tree snapshot of the current screen: every visible element with its text, role and a |
| The screen as MCP image content, so the model sees it directly. Default: JPEG, longest side 1024 px (tens of KB). |
| Just the foreground app (package, activity) and whether a crash dialog is showing. |
| Wait until some text appears on screen, up to |
Acting on the screen
Tool | Description |
| Tap by |
| Long-press by |
| Type into a field. Can tap a |
| back, home, enter, delete, tab, escape, up, down, left, right, space, power, volup, voldown, recent, menu, wakeup, or a numeric keycode. |
| Swipe between two points in device pixels. |
| Scroll a container up/down/left/right. Pass the |
Apps
Tool | Description |
| Launch an installed app by package name (e.g. |
| Open an http/https/market URL, optionally in a specific app. |
| Start an activity with any intent action, e.g. |
| Force-stop an app. |
| Installed apps; |
| Upload and install an APK from a file on the machine running this server. Returns package name, version, and whether the install succeeded. |
| Uninstall an app by package name. |
| The device's crash log buffer ( |
Quick example
Once the server is connected, ask your agent in plain language, e.g. "List my ERIUS devices, go to the home screen, and open Contacts." Here's what happens under the hood. The outputs below are real, lightly trimmed.
1. Find your device. The agent calls list_devices:
{"devices": [
{"id": "vanilla", "description": "redroid Android 13, no Google services", "state": "device",
"bootCompleted": true, "android": "13", "sdk": 33, "screen": {"width": 720, "height": 1280}},
{"id": "tablet1", "form": "tablet", "description": "redroid Android 13 (vanilla, tablet)", "state": "device",
"bootCompleted": true, "android": "13", "sdk": 33, "screen": {"width": 1600, "height": 2560}}
]}2. Go home and look at the screen. press_key(key="home", device="vanilla")
returns {"ok": true}, then snapshot(device="vanilla") returns the
foreground app plus this tree:
- Group
- ScrollView [ref=1]
- Group
- AppWidgetHostView (Search)
...
- Text [ref=4] "Gallery" (Gallery)
- View (Home)
- Group
- Text [ref=5] "Contacts" (Contacts)
- Text [ref=6] "WebView Browser Tester" (WebView Browser Tester)
- Text [ref=7] "Camera" (Camera)3. Tap something. tap(device="vanilla", ref="5"), tap(device="vanilla", text="Contacts") and tap(device="vanilla", x=102, y=1116) all open Contacts
and return {"ok": true}. After any action, take a fresh snapshot, because
refs are only valid until the next one.
4. Check the result. foreground(device="vanilla") now reports the
Contacts app, and screenshot(device="vanilla") returns the screen as an
image your agent sees directly.
Installing an APK for QA works the same way:
install_apk(device="vanilla", apk_path="/path/to/app.apk")
→ {"ok": true, "package": "com.erius.dashtest", "versionName": "1.0", "label": "ERIUS Dash Test",
"adbOutput": "Performing Streamed Install\nSuccess", "ms": 94, ...}Tips and limits
Errors are passed through. If the API refuses an action, the tool call fails with the API's error code and message, e.g.
HTTP 404: SelectorNotFound ... nothing on screen matches {"text": "..."}orHTTP 422: no_launcher_activity. The agent can read the message and adjust.Typing needs a ready field. Keystrokes sent while a field is still opening are lost. If tapping a field opens a new screen (as Android Settings' search does), tap it, wait with
snapshot/wait_for_text, then calltype_textwithout a target. Check the result withsnapshot.ASCII-only typing.
type_textsupports printable ASCII only (an adb limitation): no emoji, accented characters or non-Latin scripts.Scroll by container.
scrollneeds the ref of the scrollable container (aScrollView/Listline in the snapshot). If the container has no ref, useswipe.Screenshots cost context. The default screenshot is a JPEG with its longest side at 1024 px (tens of KB). A full-size PNG of a phone (1440×3120) is about 3 MB, more than many MCP clients accept in one tool result. Prefer
snapshot(text, much smaller) and take screenshots only when you need pixels.Coordinates from a scaled screenshot.
x,yare device pixels unless you passimage_widthandimage_height(the size of the screenshot you read them from) totap,long_pressorswipe.install_apkreads a local file. The path is on the machine runningerius-phone-mcp, not on the phone. The APK is streamed from disk and uploaded in full on each call.One server process = one API key. To use several accounts, add several server entries with different names and keys.
Rate limits. The API allows about 30 requests per second per IP and queues at most 8 requests per device; beyond that you get HTTP 429.
Troubleshooting
ERIUS_PHONE_API_KEY is not set: the variable didn't reach the server process. Put it in the client'senvblock, not just your shell. The server also prints a warning about this on stderr at startup, which shows up in your client's MCP logs.no device id given and ERIUS_PHONE_DEFAULT_DEVICE is not set: passdevice=in the tool call or setERIUS_PHONE_DEFAULT_DEVICE.list_devicesshows your ids.HTTP 401: the key is wrong or revoked.HTTP 403: the device id doesn't exist or isn't on your account. Runwhoamito see which devices your key can use.HTTP 503 device_offline: the device is restarting. Retry shortly.could not reach the ERIUS PHONE API at ...: the URL is wrong, or the API is unreachable from your machine. CheckERIUS_PHONE_API_URL, and raiseERIUS_PHONE_TIMEOUTif you're on a slow link.
License
MIT. The full license text is in the LICENSE file shipped with the package.
Available Tools
21 toolsforegroundForegroundA
Get just the foreground app: package, activity, and whether a crash dialog is showing. Cheaper than a full snapshot when you only need this.
Inputs: device. Returns: object
get_crash_logGet Crash LogA
Read the device's crash log buffer (logcat -b crash) — use this after an app crashes during QA testing to see the real stack trace.
Inputs: lines, device. Returns: object
get_deviceGet DeviceA
Get full status of one device: boot state, Android version, screen size,
and the app currently in the foreground. Omit device to use the default.
Inputs: device. Returns: object
install_apkInstall ApkA
Install an APK from a local file path on the machine running this MCP server (not on the phone) onto the device. Use this for APK QA testing — the real use case ERIUS PHONE is built for. Returns the parsed package info (name, version) and whether the install actually succeeded.
Inputs: apk_path (required), device, allow_downgrade, grant_permissions. Returns: object
launch_appLaunch AppA
Launch an installed app by package name (e.g. com.android.chrome).
Pass activity only if the app has no launcher activity resolvable automatically.
Inputs: package (required), device, activity. Returns: object
list_appsList AppsB
List installed apps on the device. Set third_party_only=True to skip pre-installed system apps and see only user-installed ones.
Inputs: device, third_party_only. Returns: object
list_devicesList DevicesA
List every Android device your API key can access, with live status (booted, Android version, screen size). Call this first to get device ids.
Inputs: no named parameters. Returns: object
long_pressLong PressA
Long-press an element (e.g. to open a context menu or start a drag).
Targets work like tap(): ref, text/content_desc, or raw x,y
(with image_width/image_height when they come from a scaled screenshot).
ms (hold time, default 800) only applies to the x,y form.
Inputs: x, y, ms, ref, +5 more. Returns: object
open_urlOpen UrlC
Open a URL (http/https/market) on the device, optionally forcing it into a specific app package (e.g. open a link in Chrome specifically).
Inputs: url (required), device, package. Returns: object
press_keyPress KeyB
Press a hardware/navigation key: back, home, enter, delete, tab, escape, up, down, left, right, space, power, volup, voldown, recent, menu, wakeup.
Inputs: key (required), device. Returns: object
screenshotScreenshotA
Take a screenshot of the device screen right now and return it as an image you can look at, plus one line of text with the image size and the device's real screen size.
The default is a JPEG whose longest side is 1024 px (tens of KB). A full-size PNG of a 1440x3120 phone is about 3 MB, which is more than many MCP clients accept in a tool result, so ask for it only when you need it: format="png", max_side=None.
format is "jpeg", "webp" or "png"; quality is 1-100 (jpeg/webp);
max_side caps the longest side in pixels (64-8192); scale (0.1-1)
shrinks by a factor instead. When the image is smaller than the screen,
pass its size as image_width/image_height to tap(), long_press() or
swipe() so x,y are read in the image's pixels. Prefer snapshot() when you
only need text and refs: it's far smaller.
Inputs: scale, device, format, quality, +1 more.
scrollScrollA
Scroll a list/container up, down, left, or right. ref must be the
ref of the scrollable container itself (a ScrollView or List line in the
snapshot), not of a row inside it: scrolling a single row does nothing.
If the container has no ref, use swipe() instead.
Inputs: ref (required), direction (required), device. Returns: object
send_intentSend IntentB
Start an activity with an arbitrary Android intent (am start -a ACTION),
e.g. action="android.settings.WIFI_SETTINGS" to jump straight to a settings
page. component is "package/.Activity"; extras values become string,
int or boolean extras by their type. Returns ok plus the new foreground app.
Inputs: action (required), data, device, extras, +2 more. Returns: object
snapshotSnapshotA
Get an accessibility-tree snapshot of the current screen: every visible
element with its text, role, and a ref id you can pass to tap/type/scroll.
This is how the agent 'sees' the phone — read it before acting, and take a
fresh one after any action since refs reset on every snapshot.
Inputs: device, max_depth, max_nodes. Returns: object
stop_appStop AppB
Force-stop a running app by package name.
Inputs: package (required), device. Returns: object
swipeSwipeB
Swipe from one point to another: in device pixel coordinates, or, with
image_width and image_height, in the pixels of a screenshot of that size.
Inputs: x1 (required), x2 (required), y1 (required), y2 (required), +4 more. Returns: object
tapTapA
Tap an element. Prefer ref from a snapshot(), or text/content_desc
to tap by visible label. Falls back to raw x,y pixel coordinates: in
device pixels, or, with image_width and image_height, in the pixels
of a screenshot of that size (the API maps the point to the screen).
Inputs: x, y, ref, text, +4 more. Returns: object
type_textType TextA
Type text into a field. Optionally tap a ref or text_target label
first to focus the field, clear the existing value, or press Enter after.
Keystrokes sent while the field is still opening are lost, so if tapping
the field opens a new screen (e.g. a search page), tap it first, wait for
it with snapshot() or wait_for_text(), then type with no target. Check the
result with snapshot(). Only printable ASCII is supported (adb limitation).
Inputs: text (required), ref, clear, enter, +2 more. Returns: object
uninstall_appUninstall AppC
Uninstall an app by package name.
Inputs: package (required), device. Returns: object
wait_for_textWait For TextB
Wait until text appears somewhere on screen (e.g. after a slow screen
transition), up to timeout_ms (max 120000). Returns the settled snapshot.
Inputs: text (required), device, timeout_ms. Returns: object
whoamiWhoamiA
Show which ERIUS PHONE account (tenant) and key id this server is using, and the device ids it can access. Handy to check the API key works.
Inputs: no named parameters. Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 21 tools
Most tools map to clearly distinct device actions, and the descriptions proactively explain when to use alternatives (e.g. snapshot vs screenshot vs foreground, tap vs swipe vs scroll). Minor overlap exists: whoami and list_devices both surface device ids, and launch_app, open_url, and send_intent can all start an app/activity. Descriptions mostly resolve these, so an agent can usually pick the right one.
All names are snake_case with a consistent convention and no style mixing. Most have a clear verb_noun form (get_device, launch_app, install_apk, wait_for_text), though several core interaction tools are bare nouns (snapshot, screenshot, foreground) or verbs (tap, swipe, scroll), which is a slight deviation.
21 tools is on the heavier end but appropriate for a full device-automation/QA surface. Each tool corresponds to a real, distinct capability (input gestures, app lifecycle, inspection, crash logs). A few could theoretically be consolidated (whoami into list_devices) but nothing feels wasted.
Coverage is strong: device discovery, screen reading (snapshot/screenshot/foreground), all input gestures, app launch/install/uninstall, intents, URL opening, and crash logs are present. Minor gaps remain—no general logcat capture beyond crash buffer, and no file push/pull or clipboard for richer QA workflows.
Maintenance
Related MCP Connectors
Cloud Android phones for AI agents: create a phone, read the screen, tap, type, install APKs, park.
MCP server for AI dialogue using various LLM models via AceDataCloud
Make games with AI: one MCP server that connects your agent to Unity, Unreal and Godot editors.
Give your AI agent a real cloud Android phone: push APKs, drive it, get screenshots and logs.
Related MCP Servers
- AlicenseAqualityCmaintenanceMCP server that gives AI agents full vision and control over Android devices via ADB and scrcpy. Supports screenshots, input, apps, UI automation, shell, files, and clipboard.38392 npm136MIT
- AlicenseNot gradedqualityCmaintenanceAn MCP server that lets AI agents control iOS and Android devices (tap, scroll, type, take screenshots, read UI trees, and run code). Works with multiple devices at the same time.94 npm54MIT
- AlicenseBqualityDmaintenanceAn MCP server that allows AI agents to drive real Android devices via adb, capturing screenshots, reading the live UI tree, and performing actions like tap, swipe, and type.135 npm1Apache 2.0
- AlicenseAqualityDmaintenanceAn MCP server that gives AI agents full control of Android devices and emulators through plain ADB — no companion APK, no extra daemon, no telemetry.2614 npm12MIT