Marionette Flutter 通过 MCP 协议使 AI 代理能够与 Flutter 应用交互,允许 Claude、Copilot 和 Cursor 等工具执行点击、滚动、输入和截屏等操作。它为 Flutter 应用扩展了自动化测试和 AI 驱动的 UI 交互功能,提升了开发工作流效率。该包可无缝集成到 Android、iOS、Web、macOS、Windows 和 Linux 等多个平台。专为在应用测试或自动化中利用大语言模型(LLM)的开发者设计,它弥合了 AI 代理与原生应用行为之间的差距。
适合场景
使用AI代理的自动化UI测试
AI 驱动的应用程序探索和调试
与 Copilot 和 Claude 的集成以实现应用程序交互
测试期间的实时截图捕获
通过MCP实现跨平台应用程序交互
主要特点
MCP协议支持AI代理通信
支持点击、滚动、输入和截图操作
多平台兼容性(Android、iOS、Web、macOS、Windows、Linux)
与基于大语言模型的开发工具无缝集成
轻量级 Flutter 扩展用于代理交互
适合对象
使用AI代理进行测试的开发者
实施AI驱动的UI自动化团队
需要高级测试编排的Flutter项目
与 Copilot、Claude 或 Cursor 合作的集成商
专注于应用开发中生成式AI的项目
使用建议
需要具有 MCP 支持的兼容 AI 代理
可能需要额外设置以实现安全的代理访问
仅限支持 MCP 的平台
依赖外部AI工具以实现完整功能
不适用于生产环境中的用户交互
原始 README
以下为关联源码仓库的 README 原文快照,最新内容请访问源码仓库。
Marionette MCP
"Playwright MCP/Cursor Browser, but for Flutter apps"
Marionette MCP enables AI agents (Claude Code, Copilot, Cursor, Gemini CLI, and more) to inspect and interact with running Flutter applications. It connects your AI agent directly to a running app via the Model Context Protocol (MCP), so it can see the widget tree, tap elements, enter text, scroll, and capture screenshots for automated AI-driven smoke testing and interaction.
Marionette MCP keeps the surface area intentionally small. It exposes only a handful of high-signal actions and returns the minimum actionable data, which helps keep prompts focused and context sizes under control.
The official Dart & Flutter MCP server focuses on development-time tasks: searching pub.dev, managing dependencies, analyzing code, and inspecting runtime errors. It can also drive the UI, but it does so through Flutter Driver, which introduces extra instrumentation in your app. Marionette MCP focuses solely (and in an opinionated way) on runtime interaction: tapping buttons, entering text, scrolling, and taking screenshots, while requiring minimal changes to your app. Use Flutter MCP to build your app, use Marionette MCP to test and interact with it with minimal code changes.
Quick Start
[!NOTE]
Your Flutter app must be prepared to be compatible with Marionette. The steps below are the short version — the Getting Started guide walks through each one.
Prepare your app — add marionette_flutter and initialize MarionetteBinding in main.dart:
Run your app in debug mode — flutter run, then copy the VM service URI from the console (e.g. ws://127.0.0.1:9101/ws).
Connect and interact — ask your agent to connect using that URI and start driving the app.
[!IMPORTANT]
Standard Material widgets work out of the box. If your app uses a custom design system, configuration is required — otherwise the agent can't see or tap your custom buttons and fields. Start with the Production Setup Checklist.
What you can do
Once connected, an agent can drive your app with a small, focused toolset: inspect the widget tree (get_interactive_elements), tap / secondary_tap / double_tap / long_press / swipe / pinch_zoom / scroll_to, enter_text, press_back_button, take_screenshots, read get_logs, and hot_reload. Full list: MCP Tools.
Your app can also expose its own actions to the agent via Custom Extensions — navigate by route name, seed test data, toggle feature flags, and more.
Some real-world prompts:
"I implemented the Forgot Password screen — connect, navigate to login, tap 'Forgot Password', enter a valid email, submit, and check the logs that the API call fired."
"I refactored the routing. Run a smoke test: cycle through all bottom-nav tabs and verify each screen loads without exceptions in the logs."
"Investigate the unresponsive 'Clear Cache' button on Settings — find it via get_interactive_elements, tap it, and analyze the logs."