Finalrun: Spec-driven mobile testing with English YAML, vision, and LLMs

What it is

It has been reported that finalrun-agent is an AI-driven CLI that lets you write human-readable test specs in YAML and run them against Android and iOS targets. Think plain-English steps — launch the app, enter credentials, verify the home screen — but powered by language and vision models under the hood. The project’s README shows an installer you can curl-and-bash to set up Node, the CLI, model “skills” (allegedly for Claude Code / Codex), and Android/iOS tooling. There’s even a demo on YouTube if you prefer watching to reading.

How it works

You create a .finalrun workspace in your repo, drop YAML tests under .finalrun/tests, configure an AI provider key and environment (.env files live at the workspace root), and run finalrun check/test/doctor to validate and execute runs. Artifacts and native driver assets are kept under ~/.finalrun/, and the tool exposes commands to list runs or serve reports locally. The README highlights model examples (google/gemini-3-flash-preview) and warns: stash secrets only in workspace-root .env files — not in the YAML specs. Practical stuff. No magic. Just careful wiring.

Why it matters

Flaky UI tests are the bane of mobile devs. Could an LLM + vision approach make them less painful? Possibly. This sits squarely in the trend of “AI for developer workflows” — Copilot for QA, if you will — and it promises faster iteration for teams who want spec-driven tests written in plain language. Will it replace handcrafted instrumentation or device farms? Not overnight. But if you’ve ever sighed at another brittle test, Finalrun’s clean UX and emphasis on workspace hygiene might feel like a breath of fresh air. Want to try? The repo and install script are on GitHub; a cloud device waitlist and Slack community are linked from the project page.

Sources: github.com/final-run, Hacker News