[sozal]

Yeah, you’re right. The most of the gain is from the execute tool (batching many steps in one call), not from CDP vs Playwright. If Playwright MCP added a similar sandboxed “run script that calls other tools” and the model used it the same way, token use and turn count would likely get much closer. So the win is API design (batching + scoping + lean tool defs), not the protocol. The post is about showing that this kind of design matters, we’re not claiming our stack is inherently lighter.