AI device automation and executable-skill risks

Open-source ClickClickClick uses large-model vision to automate Android and macOS interactions without coordinate scripts.
Key insights
ClickClickClick enables natural-language device automation
ClickClickClick is an open-source project that leverages large models' visual capabilities to automate Android devices and macOS machines. It avoids rigid coordinate-based scripting by accepting natural-language commands like “open Gmail and send an email” or “find bus stops on the map,” and the AI autonomously completes UI interactions. The project is compatible with OpenAI, Claude, Gemini, and can use local models via Ollama, and requires ADB configuration to control Android devices.
Use cases include app testing and routine task automation
The ClickClickClick project is presented as a framework for automating app testing, software interaction flows, and repetitive daily tasks, offering a practical experiment platform to delegate UI interactions to AI without manual scripting.
Executable Skills and scripts carry security risks
A cautionary post states Skills' main value often lies in their accompanying scripts and tools rather than prompts alone, and if those scripts come from untrusted providers they become a vector for malicious actions. The post advises against using third-party Skills that include executable scripts and warns that poorly tested scripts—even those written by oneself—can accidentally delete important files.