The video opens with the current AI landscape: the host notes strong buzz around a GPT-6 Astra model from OpenAI that, in his experience, pulls ahead of Claude in places. He immediately balances that with Claude-side progress, especially browser integration, which is the video's real subject. His framing is that two races run in parallel — raw model quality and everyday tool integration — and both deserve attention.

The old limitation is worth recalling: Claude could already visit sites and fetch listings, but it stopped at logins, button clicks, and form inputs. The video's claim is that an early-September update removed that wall: Claude Code can now drive the user's own Chrome when permission is granted, or fall back to an anonymous profile otherwise.

Setup is shown step by step. First the Claude desktop app is installed, with the note that at least the 20-dollar plan is required. Then, instead of plain chat, the host switches to Cowork — the workspace for file creation, reading, and reporting. Browser power is enabled from profile settings via a 'Claude in Chrome' toggle at the bottom, with per-site choices such as ask every time, allow once, or block entirely.

The first live test uses an everyday scenario: find a comfortable keyboard from a given brand, visit the store, add it to the cart. After the instruction is entered in Cowork, Claude asks clarifying questions about keyboard type and budget, then surfaces a permission dialog. Once approved, a separate tab marked with a Claude badge opens, and the model's browsing, product discovery, and cart navigation can be watched step by step.

Two details stand out in this demo. First, the safety balance: Claude pauses before any purchase confirmation, because the connected profile is the real account and one click could place the order. Second, the reporting quality: the finished task comes back with the added product, an alternative price band, a product link, and even a note like 'no wrist rest included, want one?' — an assistant that compiles results, not just clicks.

The second scenario is job hunting: backend listings focused on .NET and C# with a few years of experience, searched on the X platform. Claude again clarifies criteria first, adjusts the experience range, asks for region, then opens the job board and applies filters. The task keeps running in the background while the video continues, and the host notes the output can be turned into a Word document and emailed via the Gmail connector. Chained jobs — find a free AI tool, download the repo, get the project running — are described as following the same logic.

The second half covers a separate capability: the Chrome extension. Installation is a one-click add from the Web Store, the icon gets pinned to the toolbar, and after sign-in a side panel opens on every page. A banner notes the page is being reviewed by Claude; anything on that page can then be summarized or queried. The host's distinction is clear: the Cowork side is more capable with device and connector access, while the extension is the more practical interface for non-file, non-code work.

The first extension experiment runs on the host's own video page: scan the comments, hit like if positive comments dominate and dislike if negative ones do, then draft a reply without sending it. Claude screenshots the page, reviews comments one by one, casts the like, and leaves the draft in the comment box. The send step is deliberately left to the human.

The second extension test is more playful: the host opens a browser game he built himself and instructs Claude to play the single-player mode and report defects. Claude actually plays, navigates the screen, and hunts for issues. No major bug surfaces — the host credits earlier self-testing — but the idea of running an AI as a game tester is concretely demonstrated.

The closing picture is about parallel work: job listings keep being gathered on the Cowork side while comment analysis and game testing run simultaneously on the extension side, three or four jobs at once. Extension work also lands in chat history, so there is no confusion about what was asked where. The practical tip at the end: the panel can be closed or the feature disabled from settings when done. The host's verdict is direct: anyone wanting to use Claude with Chrome productively should try both capabilities.

AI commentary

"To me, this video captures where the AI race is really heading: raw model smarts matter, but deep penetration into everyday tools matters just as much. Having Claude click and browse in my place sounds like a small convenience, yet it turns everything from shopping to job hunting into something I can delegate."

Sources

claude · chrome · cowork · browser agent · automation