FinalRun just dropped an open-source CLI that tests mobile apps using vision AI instead of brittle selectors. You write tests in plain YAML, the agent "sees" your app like a user would, and it adapts when your UI changes. The tool supports OpenAI, Google, and Anthropic models, with Google's Gemini 3 Flash Preview as the default.
The pitch is straightforward: traditional mobile testing tools like Appium break constantly because they rely on DOM selectors and XPath locators that shift whenever someone moves a button. FinalRun claims roughly 99% reduction in flaky tests by using computer vision to identify elements visually.
Early adopters report jumping from 40% test coverage with Appium to full coverage within weeks, according to the project's documentation.
What makes this genuinely **agentic** is how it handles the whole QA workflow. The system can run targeted tests based on code changes in each commit. It detects issues like button overlaps and API errors through log analysis. And it files tickets automatically to Jira or Linear with video evidence attached. It also integrates with Claude Code through 'skills' so your AI coding assistant can trigger relevant tests as you work.
The business model is open-core. The CLI is free and Apache-licensed, while the company sells cloud device infrastructure and enterprise CI/CD integrations. Installation pulls down Node.js, platform tools, and the CLI in one script.
If the 99% flakiness reduction holds up across large test suites, maintaining mobile tests might stop being a nightmare. If vision-based testing degrades on complex UIs or unusual screen densities, teams will find out fast.