One Click in the Browser, Context for Any Agent

The submission hit Hacker News on 20 July 2026 with one point and zero comments, so expectations should be calibrated accordingly: this is not a launch with traction, it is a personal fix for a very recognisable piece of friction. The author explains it plainly in his Medium article: VS Code already lets you add page elements to the Copilot chat, but that capability lives inside Copilot. He wanted the same thing for any agent, so he built an extension that lets you click a web page element and paste it into whichever AI chat you happen to be using.

Anyone who has debugged an interface with an assistant knows the scene. You describe the broken component by hand, paste an HTML fragment pulled from the inspector, attach a screenshot, and the model ends up answering about a different button. The context exists and it is right there in your browser, but the route to the chat window is manual. Every copy and paste loses something: the exact selector, the Tailwind class, the state the component was breaking in.

Context is becoming the lock-in point

The interesting part is not the extension, it is what it points at. For two years the competition between coding assistants was mostly played out on the model. Now it is increasingly played out on context capture: what the agent knows about your repository, your terminal and your browser without you having to tell it. That layer is what ties a user to a specific product, because switching assistants no longer means switching models, it means rebuilding all the context plumbing you had set up.

An agent agnostic extension breaks that pattern cheaply: the editor generates the context and the person writing decides where it goes. For anyone alternating between Claude Code in the terminal and another assistant inside the editor, that neutrality has immediate practical value, not just ideological value.

Where MCP fits and where it does not

Two things that often get mixed up are worth separating. A browser MCP server, along the lines of the Playwright one, gives the agent autonomous access: it navigates, inspects the DOM and acts on its own. What this article describes is the opposite in terms of control: the person points and the agent receives. They are complementary approaches, not rivals. The first scales repetitive tasks and test flows; the second solves the case where you already know which element is broken and want the model to know it too without writing twenty lines of description.

For teams already working with skills and subagents, the pattern fits as a preliminary step. First you narrow the context by hand, then you let the agent do the heavy lifting from that anchor. The more precise the starting point, the less drift there is in the answer.

Who it is useful for

Front-end and QA people who review interfaces in a browser most of the day.
Design and product profiles who report issues and would rather not wrestle with the inspector.
* Teams that alternate between several assistants and do not want to rebuild the workflow in each one.

It adds little if you work on the backend, if your review loop already runs through automated screenshots in continuous integration, or if your agent has direct browser access via MCP and you trust it to locate the element on its own.

What we still do not know

The article is from today and the thread on Hacker News is empty, so there is no independent review and no third party experience. Reasonable questions remain before installing anything: what permissions the extension requests over the pages you open, whether captured content stays local, how it behaves with applications behind authentication, and who will maintain it six months from now. We have tried plenty of utilities like this one and the mortality rate is high.

Our reading is measured and favourable. The idea is worth more than this particular implementation, and the likely outcome is that editors absorb it themselves or an MCP server ends up doing the same thing in a standard way. In the meantime, anything that shortens the distance between what you see on screen and what the model understands is worth five minutes of testing.