Example prompt:

“Buy milk from Blinkit”

Claude then:

• searches products • adds them to the cart • selects the delivery address • proceeds to checkout

Under the hood the flow looks like this:

Claude Desktop → MCP tool call → Python MCP server → Playwright browser automation → Blinkit website

Instead of relying on APIs, the agent interacts with Blinkit by controlling a real browser session through Playwright.

So Claude effectively becomes an agent that can complete a real world shopping task. The project started as an experiment around a simple question: If LLMs can control browsers reliably, do we still need service APIs for automation? Curious what other MCP integrations people are experimenting with.