Most AI demos stop at chat. But what happens when AI can actually open websites, click buttons, compare options, create files and download invoices for you? That's Claude working with Playwright MCP, a plugin that lets Claude control a real browser: click buttons, fill forms, extract data from pages.

Claude has its own browser use for quick autonomous lookups; Playwright shines for structured, repetitive workflows, like extracting your invoices from the same portals every month. Installing it is one prompt: "go install this MCP locally."

Demo 1: 60+ car listings into a ranked Excel table

Shopping on mobile.de for a BMW 5 Series under €65K: filters get you to 62 offers, but comparing the extras packages across them is where the real time goes, a pricier car can be the better deal once you account for what's included.

I gave Claude the results URL and asked for an Excel comparison: brand, model, year, price, mileage, seller, warranty, interior, extras, plus its own price-quality verdict per car. It launched multiple agents in parallel to work through the listings and came back ~30 minutes later with a ranked spreadsheet, scored from "good" to "no go", with reasoning and links. I now review a top-10 instead of 62 tabs.

Demo 2: Publishing a WordPress blog post

I handed Claude a written blog post and asked it to publish it on my WordPress site via the browser, logging in myself when prompted. It navigated to the editor, generated a cover image with GPT-Image, inserted the title, headings, hyperlinks and body, picked the right categories, and hit publish. (It skipped tags, because I didn't ask. Prompt specificity matters.)

Security tip: don't hand AI your admin credentials. Create a limited user that can only draft and edit posts, and give it that login.

Demo 3: Downloading invoices for accounting

Every month I collect invoices from services my company subscribes to. I asked Claude to pull my ChatGPT and Claude invoices into an "April 2026" folder. It navigated the billing pages, waited for me to log in, and, after one correction (it first saved page screenshots instead of pressing the actual download button), saved the real PDF invoices.

Honest lessons from the trenches

  • Be step-by-step for recurring workflows. Zero-shot "go figure it out" works but is slow; telling it exactly where to click is what Playwright was built for. Bonus trick: after a successful run, ask Claude to write down the step-by-step instructions it just followed, that becomes your reusable prompt.
  • It sometimes gets stuck, stop and say "continue", and it picks right up.
  • Long browser sessions bloat the context (every step includes a screenshot), and quality degrades as context grows. Keep tasks scoped.
  • Credentials are your risk. You're sharing confidential access with a non-deterministic system on third-party sites, use view-only accounts wherever possible.

If your company has repetitive browser work, reporting, finance processes, content workflows, this is exactly the kind of automation we build at FlowByte.