leancodepl/marionette_mcp

What it solves

Marionette MCP allows AI agents to interact with and test running Flutter applications in real-time. It bridges the gap between an AI's reasoning and a live app's UI, enabling agents to perform automated smoke testing, investigate bugs, and verify feature implementations without requiring the developer to manually drive the app.

How it works

It uses the Model Context Protocol (MCP) to connect AI agents (such as Claude Code, Cursor, or Copilot) to a Flutter app running in debug mode. By adding a binding to the Flutter app, the system exposes a focused set of tools that allow the agent to inspect the widget tree, perform gestures (tap, swipe, scroll), enter text, capture screenshots, and read application logs.

Who it’s for

Flutter developers and QA engineers who want to use AI agents to automate the interaction and testing of their mobile or desktop applications.

Highlights

• Runtime Interaction: Focuses on driving the live UI rather than development-time code analysis.
• Focused Toolset: Provides high-signal actions (tap, scroll, enter text) to keep AI context windows small and prompts efficient.
• Extensible: Supports custom extensions to expose app-specific actions (like seeding test data or toggling feature flags) to the agent.
• Broad Agent Support: Compatible with various AI tools including Claude Code, Gemini CLI, and Cursor.