What it is
A Chrome extension that inverts the usual browser agent model: instead of describing a goal in a prompt, the user selects elements on the live page and describes the desired outcome. It runs BYOK, sandboxed, and returns an editable office file.
Why it showed up
2.9k stars. Selection-first is a genuinely different interaction model from prompt-first browser agents, and it sidesteps the reliability problem of asking a model to guess the right selector.
Who should look at it?
People who work with messy webpages and want to keep a human in the loop should find the interaction model easy to understand. Selecting the relevant page region first narrows the task before the agent has to interpret the request.
What to inspect
Pay attention to the boundary between the Chrome extension, the selected page content and the generated office file. The important implementation questions are what leaves the browser, how BYOK credentials are handled, and whether the output remains editable after the agent finishes.
Watch point
Selection-first reduces ambiguity; it does not remove the security and reliability concerns of browser automation. Treat the sandbox and credential model as things to verify in the repository before using it with real sessions.