Ah I had two motivations, it started when ralphs and amazon were blocking my agent, it's then pretty easy to ask the agent to solve their own problem if you give it compute resources.

I also wanted to play with some ideas I have around approval of agentic actions and sharing of credentials. I can selectively share browser cookies with my MCP server and it'll inject them into browsers the agent can use for a limited scoped task. I use a decision API to try and see if the browser actions match the approved work or if they're going awry. This is generally preferable (for me) vs just handling my credentials over and letting the agents YOLO with their cloud browser. I get saved recordings sent to my phone.