AI E2E Testing — 8 Platforms, Zero Test Code

flutter-skill is an MCP server that connects AI agents to running apps across 8 platforms. The agent takes screenshots, taps elements, enters text, scrolls, navigates, inspects UI trees, and verifies state — all through natural language.

Supported Platforms

Install

# npm (recommended)
npm install -g flutter-skill

# Homebrew
brew install ai-dashboad/flutter-skill/flutter-skill

# Or download binary from GitHub Releases

MCP Configuration

Add to your AI agent's MCP config (Claude Desktop, Cursor, Windsurf, OpenClaw, etc.):

{
  "mcpServers": {
    "flutter-skill": {
      "command": "flutter-skill",
      "args": ["server"]
    }
  }
}

Quick Start

1. Initialize your app (one-time)

flutter-skill init

Auto-detects project type and patches your app with the testing bridge.

Verify: Output should confirm the project type was detected and main entry point was patched. If it fails, check that you are in the project root and the framework is supported.

2. Launch and connect

flutter-skill launch .

Verify: A VM Service URI appears in the output (e.g. ws://127.0.0.1:50000/ws). If no URI appears, check that Flutter/the target framework is installed and the app compiles.

3. Test with natural language

The agent follows this core loop:

1. screenshot() — see the current screen
2. inspect_interactive() — discover all tappable/typeable elements with semantic refs
3. tap(ref: "button:Login") — tap using stable semantic reference
4. enter_text(ref: "input:Email", text: "[email protected]") — type into field
5. waitforelement(key: "Dashboard") — verify navigation succeeded
6. screenshot() — confirm final state

If inspectinteractive() returns no elements: Take a screenshot to confirm the screen loaded, then check getlogs() for errors. The app may still be loading — retry after a short wait.

Available MCP Tools

Core Actions

Inspection

Text Manipulation

App Control

Semantic Refs

inspect_interactive returns elements with stable semantic reference IDs:

button:Login          → Login button
input:Email           → Email text field
toggle:Dark Mode      → Dark mode switch
button:Submit[1]      → Second Submit button (disambiguated)

Format: {role}:{content}[{index}]

7 roles: button, input, toggle, slider, select, link, item

Use refs for reliable element targeting that survives UI changes:

tap(ref: "button:Login")
enter_text(ref: "input:Email", text: "[email protected]")

Testing Workflow

Core Loop

screenshot() → inspect_interactive() → tap/enter_text → screenshot() → verify

Always call screenshot() before and after actions. Use waitforelement() after navigation — apps need time to transition.

Validation Checkpoints

• After screenshot(): Confirm the expected screen is visible before acting.
• After tap() or enter_text(): Call screenshot() to verify the UI responded.
• After navigation: Use waitforelement(key: "targetscreen") with a timeout. If it times out, call screenshot() and getlogs() to diagnose.
• On unexpected state: Call getlogs() and inspectinteractive() to understand what elements are present.

Element Targeting Priority

1. ref: (most reliable) — semantic ref from inspect_interactive()
2. key: — widget key set by the developer
3. text: — visible text content (fragile if text changes)
4. Coordinates — last resort, breaks on different screen sizes

Links

• GitHub
• npm
• Documentation
• pub.dev
• VSCode Extension