Midscene

Every Selenium or Playwright suite carries the same hidden liability: it is bound to selectors that break the moment a designer renames a class or restructures a div. Midscene removes that coupling entirely — it reads the screen the way a human does, from pixels, so a test describes intent ("click the login button") rather than structure ("div.header > button:nth-child(2)").

What Sets It Apart

• Vision-first, not DOM-first — it acts on screenshots, so one script runs against a web page, an Android app, iOS, HarmonyOS, or a desktop window without rewriting locators.
• Natural-language primitives — aiAct performs actions, aiQuery extracts structured data, and aiAssert verifies visual state, all from plain-English instructions.
• Bring your own model — it is a harness, not a model: plug in Qwen, Doubao, GLM, or Gemini, or self-host UI-TARS, trading cost, latency, and privacy as you see fit.
• Two modes — drop it into an existing Playwright/Vitest pipeline, or let it run as an autonomous agent that works out the steps itself.

Who It Is For

A strong fit for teams whose UIs change fast, who test across several platforms, or who want non-engineers to author tests in prose. The tradeoff is real: every step is a model inference, so runs are slower and cost more per action than selector-based tools, and visual ambiguity can still mislead the model. If your app is a stable single web surface with a mature selector suite, classic Playwright stays cheaper and more deterministic.