UI Automation
Control and automate your Tauri application's UI. Webview tools work across Linux, Windows, and macOS. Native dialog automation is Windows-only.
Multi-Window Support
All webview tools support targeting specific windows in multi-window applications. Use the optional windowId parameter to specify which window to interact with. If not specified, tools default to the "main" window.
Discovering Windows
Use manage_window with action: "list" to discover all available windows:
{
"tool": "manage_window",
"action": "list"
}Response:
{
"windows": [
{
"label": "main",
"title": "My App",
"url": "http://localhost:1420/",
"focused": true,
"visible": true,
"isMain": true
},
{
"label": "settings",
"title": "Settings",
"url": "http://localhost:1420/settings",
"focused": false,
"visible": true,
"isMain": false
}
],
"defaultWindow": "main",
"totalCount": 2
}Getting Window Info
Use action: "info" to get detailed information about a specific window:
{
"tool": "manage_window",
"action": "info",
"windowId": "main"
}Response:
{
"width": 800,
"height": 600,
"x": 100,
"y": 100,
"title": "My App",
"focused": true,
"visible": true
}Resizing Windows
Use action: "resize" to resize a window to specific dimensions:
{
"tool": "manage_window",
"action": "resize",
"width": 1024,
"height": 768
}Response:
{
"success": true,
"windowLabel": "main",
"width": 1024,
"height": 768,
"logical": true
}By default, dimensions are in logical pixels (respects display scaling). Set logical: false for physical pixels. The resize will fail if the window has fixed size constraints or is not resizable.
Targeting a Specific Window
Add windowId to any webview tool to target a specific window:
// Execute JavaScript in the settings window
{
"tool": "webview_execute_js",
"script": "document.title",
"windowId": "settings"
}
// Take a screenshot of the main window (explicit)
{
"tool": "webview_screenshot",
"windowId": "main"
}driver_session
Manage UI automation session lifecycle. Initializes console log capture and prepares the webview for automation. Supports remote device connections via the host parameter.
Parameters
| Name | Type | Required | Description |
|---|---|---|---|
action | string | Yes | Action to perform: 'start', 'stop', or 'status' |
host | string | No | Host address to connect to (e.g., '192.168.1.100'). Falls back to MCP_BRIDGE_HOST or TAURI_DEV_HOST env vars |
port | number | No | Port to connect to (default: 9223) |
Actions
start- Start a new session, connecting to the Tauri appstop- Stop the current session and disconnectstatus- Check current connection status without changing state. Returns the app'sidentifier(bundle ID) which can be used to determine if the session is connected to the correct app
Connection Strategy
When starting a session, the tool uses the following connection strategy:
- Try localhost first - Most reliable for simulators, emulators, and desktop apps
- Fall back to configured host - If localhost fails and a remote host is configured
- Auto-discover - Scan port range on localhost for running apps
- Graceful fallback - Return success message even if no app found (allows IPC-only mode)
Example
// Start an automation session (default - localhost)
{
"tool": "driver_session",
"action": "start"
}
// Connect to a real iOS device on the network
{
"tool": "driver_session",
"action": "start",
"host": "192.168.1.100"
}
// Connect to a specific port
{
"tool": "driver_session",
"action": "start",
"port": 9225
}
// Check connection status
{
"tool": "driver_session",
"action": "status"
}Response
Start/Stop:
Session started with app: My App (localhost:9223)Status:
{
"connected": true,
"app": "My App",
"identifier": "com.example.my-app",
"host": "localhost",
"port": 9223
}The identifier field contains the app's bundle ID (e.g., com.example.my-app). Use this to verify you're connected to the correct application before reusing an existing session.
Note: The
identifierfield may benullif the Tauri app uses an older version of the MCP Bridge plugin that doesn't provide app identification. In this case, you cannot verify the app identity and should start a new session if uncertain.
Environment Variables
MCP_BRIDGE_HOST- Default host whenhostparameter not providedTAURI_DEV_HOST- Fallback host (same as Tauri CLI uses for mobile dev)MCP_BRIDGE_PORT- Default port whenportparameter not provided
Remote Device Setup
For real iOS/Android devices on the network:
- Ensure your development machine and device are on the same network
- The Tauri plugin binds to
0.0.0.0by default, allowing remote connections - Use the device's IP address as the
hostparameter
Android alternative: Use adb reverse tcp:9223 tcp:9223 to forward the port, then connect to localhost.
Note: No external driver process required.
Native Windows Dialogs
native_dialog_snapshot and native_dialog_interact automate native message, confirmation, single- or multi-file Open, folder selection, and Save dialogs. These windows are outside the webview DOM, so webview_interact, webview_keyboard, and webview_screenshot cannot inspect or control them.
The tools require an active driver_session and an interactive Windows desktop. Discovery is restricted to visible dialogs in the connected Tauri process whose bounded owner chain leads back to the targeted Tauri window. This includes nested prompts such as a Save overwrite confirmation without exposing desktop-wide automation or dialogs from other applications. Windows toast notifications are not dialogs and are not supported.
native_dialog_snapshot
Returns a bounded semantic snapshot containing control types, automation IDs, names, semantic roles, supported UI Automation patterns, owner depth, parent-dialog references, and opaque elementRef values. File-dialog snapshots include actionable navigation controls, the current-location control, navigation-tree items, and selectable file-system entries. Use windowId and appIdentifier with the same targeting behavior as webview tools. timeoutMs accepts 100–10000 milliseconds.
{
"tool": "native_dialog_snapshot",
"windowId": "main",
"minOwnerDepth": 1,
"timeoutMs": 2000
}minOwnerDepth defaults to 1, where the Tauri window directly owns the dialog. Set it to 2 to wait for a prompt owned by another app-owned dialog, such as an overwrite confirmation. Nested dialogs are returned before their parents.
Controls advertise one or more supported actions:
invoke—InvokePattern, typically buttons such as accept, cancel, yes, or nosetValue—ValuePattern, used for a complete absolute filename/pathsetPaths—ValuePattern, used for up to 100 complete absolute existing file paths in a multi-select Open dialogselect—SelectionItemPattern, where the dialog exposes a selectable control
Semantic roles are derived from automation IDs and control patterns rather than localized button text. Names and dialog text are returned to the caller for inspection but are not written to bridge logs.
native_dialog_interact
Apply one advertised action to an elementRef from the latest snapshot:
{
"tool": "native_dialog_interact",
"action": "invoke",
"elementRef": "native_..."
}For Open or Save dialogs, set the filename field to a complete absolute path, then invoke the accept button:
{
"tool": "native_dialog_interact",
"action": "setValue",
"elementRef": "native_...",
"value": "C:\\tmp\\fixture.txt"
}For a multi-file Open dialog, use the filename field's advertised setPaths action and then invoke accept:
{
"tool": "native_dialog_interact",
"action": "setPaths",
"elementRef": "native_...",
"paths": [
"C:\\tmp\\first.txt",
"C:\\tmp\\second.txt"
]
}For a folder picker, use setValue with a complete absolute directory path and invoke accept. To respond to a Save overwrite prompt, invoke Save, take a new snapshot with minOwnerDepth: 2, and invoke the confirmation's advertised affirmative action. Back, forward, up, address/current-location, navigation-tree, and file-system entry controls are included when Windows exposes their corresponding UI Automation patterns.
Element references are session-bound, expire after 30 seconds, and are otherwise ephemeral. A new snapshot replaces the previous reference set; invoking or selecting a control invalidates references for that dialog, while setValue and setPaths preserve them so the previously discovered accept control can be invoked. If the dialog changes or closes, the tool returns a stale-reference error and the caller must take another snapshot.
Supported scope includes single- and multi-file Open, Save, folder selection, overwrite confirmation, message acknowledgement, ask/confirm responses, absolute-path entry, cancellation, nested app-owned dialogs, and navigation controls that expose the patterns above. Arbitrary custom controls without UI Automation patterns, permission prompts outside the app's ownership chain, notification/toast automation, image recognition, and coordinate-based native control remain out of scope.
webview_find_element
Find UI elements using CSS, XPath, or text selectors.
Parameters
| Name | Type | Required | Description |
|---|---|---|---|
selector | string | Yes | Element selector |
strategy | string | No | Selector strategy: 'css', 'xpath', 'text' (default: 'css') |
windowId | string | No | Window label to target (defaults to 'main') |
Example
// Find a button by CSS selector
{
"tool": "webview_find_element",
"selector": "#submit-button",
"strategy": "css"
}
// Find by text content
{
"tool": "webview_find_element",
"selector": "Submit",
"strategy": "text"
}Response
Returns element information including tag name, text content, and attributes.
read_logs
Read logs from various sources: webview console logs, Android logcat, iOS simulator logs, or desktop system logs.
Parameters
| Name | Type | Required | Description |
|---|---|---|---|
source | string | Yes | Log source: 'console', 'android', 'ios', 'system' |
lines | number | No | Number of log lines to retrieve (default: 50) |
filter | string | No | Regex or keyword to filter logs |
since | string | No | ISO timestamp to filter logs since |
windowId | string | No | Window label for console logs (defaults to 'main') |
Sources
console- JavaScript console logs from the webview (requires active session)android- Android logcat outputios- iOS simulator logssystem- Desktop system logs (macOS/Linux)
Example
// Get webview console logs
{
"tool": "read_logs",
"source": "console"
}
// Get console logs matching a pattern
{
"tool": "read_logs",
"source": "console",
"filter": "error|warning"
}
// Read Android logcat
{
"tool": "read_logs",
"source": "android",
"filter": "com.myapp",
"lines": 100
}
// Read system logs
{
"tool": "read_logs",
"source": "system",
"lines": 50
}Response
Returns log entries from the specified source with timestamps and log levels.
webview_dom_snapshot
Get a structured DOM snapshot of a Tauri app's webview for AI consumption.
Parameters
| Name | Type | Required | Description |
|---|---|---|---|
type | 'accessibility' | 'structure' | Yes | Snapshot type |
selector | string | No | CSS selector to scope the snapshot |
windowId | string | No | Window label to target |
appIdentifier | string | number | No | App identifier |
Snapshot Types
accessibility - Uses aria-api for comprehensive, spec-compliant accessibility computation:
- WAI-ARIA 1.3 role computation
- Accessible names and descriptions
- ARIA states (disabled, expanded, checked, etc.)
- Best for understanding UI semantics and finding interactive elements
structure - DOM structure tree with:
- Element tag names
- Element IDs (if present)
- CSS classes (if present)
data-testidattributes (if present)- Best for understanding page layout and debugging CSS selectors
Accessibility Snapshot Format
- heading "Page Title" [level=1] [ref=e0]:
- navigation [ref=e1]:
- list [ref=e2]:
- listitem [ref=e3]:
- link "Home" [ref=e4]
- main [ref=e5]:
- button "Submit" [disabled] [ref=e6]
- textbox "Enter name" [ref=e7]Structure Snapshot Format
- body [ref=e0]:
- div#app.container [ref=e1]:
- header.header [ref=e2]:
- nav.nav-menu [ref=e3]
- main.content [ref=e4]:
- form#login-form [ref=e5] [data-testid=login]:
- input#username [ref=e6]
- button.btn.btn-primary [ref=e7]Element References
Each element includes a ref attribute (e.g., [ref=e0]) that can be used with other webview tools to target that specific element. Simply pass the ref ID as the selector parameter:
// First, get a snapshot to see available refs
{ "tool": "webview_dom_snapshot", "type": "accessibility" }
// Then use a ref to interact with an element
{ "tool": "webview_interact", "action": "click", "selector": "ref=e7" }
// Or type into an input
{ "tool": "webview_keyboard", "action": "type", "selector": "ref=e6", "text": "hello" }Refs work with: webview_interact, webview_keyboard, webview_get_styles, webview_find_element, and webview_wait_for.
WARNING
Refs are regenerated each time you call webview_dom_snapshot. If the DOM changes significantly, run a new snapshot to get updated refs.
Example
// Accessibility snapshot of entire page
{
"tool": "webview_dom_snapshot",
"type": "accessibility"
}
// Structure snapshot of a specific component
{
"tool": "webview_dom_snapshot",
"type": "structure",
"selector": ".login-form"
}Scoped Snapshots
Use the selector parameter to snapshot a subtree. If the selector matches multiple elements, each match is returned as a separate labeled snapshot.
Response
Returns a YAML-formatted tree with:
- Accessibility type: Element roles, accessible names, ARIA states, ref IDs
- Structure type: Tag names, IDs, CSS classes, data-testid attributes, ref IDs
- Metadata footer with generation timestamp and element count
webview_select_element
Activate a visual element picker overlay in the Tauri app. The user sees a blue highlight following their cursor and can click to select an element. The tool returns rich metadata about the selected element plus a cropped screenshot of it.
This is useful when you want the user to point out a specific UI element so you can discuss it, debug it, or make changes to it.
Slash Command
You can also use the /select slash command for a guided element selection workflow. See Prompts for details.
Parameters
| Name | Type | Required | Description |
|---|---|---|---|
timeout | number | No | Timeout in ms for user to pick an element (5000–120000, default: 60000) |
windowId | string | No | Window label to target (defaults to 'main') |
appIdentifier | string | number | No | App identifier for multi-app setups |
Example
// Activate the element picker with default timeout
{
"tool": "webview_select_element"
}
// Activate with a longer timeout
{
"tool": "webview_select_element",
"timeout": 120000
}Response
Returns two content items:
Text — Formatted element metadata including:
- Tag name, ID, classes
- CSS selector and XPath
- Bounding rect (position and dimensions)
- HTML attributes
- Text content (truncated to 200 characters)
- Computed CSS styles
- Parent chain (tag, id, classes, dimensions for each ancestor)
Image — A cropped PNG screenshot of just the selected element.
If the user presses Escape or clicks the X button, the tool returns a cancellation message.
How It Works
- A translucent blue overlay appears in the Tauri app
- As the user moves their cursor, elements are highlighted with a blue outline
- The user clicks an element to select it
- The overlay is removed and metadata + screenshot are returned
- The user can cancel by pressing Escape or clicking the close button
webview_get_pointed_element
Retrieve element metadata for an element the user previously pointed at via Alt+Shift+Click in the Tauri app. This is the passive counterpart to webview_select_element — the user can Alt+Shift+Click elements at any time while using the app, and then later the agent can retrieve the metadata.
Parameters
| Name | Type | Required | Description |
|---|---|---|---|
windowId | string | No | Window label to target (defaults to 'main') |
appIdentifier | string | number | No | App identifier for multi-app setups |
Example
// Get the element the user Alt+Shift+Clicked
{
"tool": "webview_get_pointed_element"
}Response
If the user has Alt+Shift+Clicked an element, returns the same content as webview_select_element:
- Text — Formatted element metadata (tag, id, classes, selector, styles, parent chain, etc.)
- Image — A cropped PNG screenshot of the element.
If no element has been pointed, returns an instruction message telling the user to Alt+Shift+Click an element first.
When to Use Which
webview_select_element— The agent initiates the picker. Best when the AI wants to ask the user to select something.webview_get_pointed_element— The user initiates the selection. Best when the user says "look at this element I pointed at".