Most computer-use demos hide the part that actually matters: the loop that turns a model's "click here" into a real browser action and feeds the result back. This repository is Google's minimal, readable reference for that loop, observe a screenshot, let the model propose a UI action, execute it, then repeat, and it tracks Gemini's current default computer-use model rather than betting on one version.
What Sets It Apart
- It exposes the agent loop, not a product. Each turn sends the task, the current screenshot, and recent action history to the model; the model replies with a UI action; the code executes it and screenshots again. Reading it teaches the control flow you would otherwise reverse-engineer.
- Tracks the live model lineup. The default is now
gemini-3.5-flash, withgemini-3-flash-previewand the earliergemini-2.5-computer-use-previewalso selectable, so the same loop keeps working as Google folds computer use into newer Gemini models. - Two interchangeable browser backends. Playwright drives a local Chrome for full control; Browserbase runs the same agent against a cloud browser. Switching is a config change, so you can move from local debugging to scaled, headless runs without rewriting the agent.
- Honest about a real failure mode. Playwright cannot capture native
<select>dropdowns on some operating systems, so dropdown-heavy forms break locally, and the README points to Browserbase as the workaround instead of hiding the issue.
Who It's For
Great fit if you want to understand or prototype browser automation with Gemini: testing UI flows, scripting multi-step web tasks, or learning how visual computer-use agents are wired before committing to a framework. Look elsewhere if you need a polished, production-ready RPA tool: this is preview-stage reference code, browser-only with no native desktop control, and you supply the safety guardrails, retries, and error handling yourself. For context, Google's dedicated 2.5 computer-use model reported roughly 70% on the Online-Mind2Web benchmark.
