I built a Jev-based wingman for browser MCP servers: it works alongside Playwright MCP, takes over the clicking, and hands back what it can't do (34/34 runs, 30% cheaper)
AI agents drive browsers one click at a time, and every click costs a full model turn. I built jev-browser-wingman to hand those clicks to TypeSafe's Jev, then benchmarked it against Claude Sonnet 5.5 doing its own clicking.
It is a wingman, not a replacement. Your browsing server (Playwright MCP in my tests) stays installed, and the agent keeps its tools for reading the page, tabs, dialogs and dragging. Wingman takes over only the clicking and typing: the agent sends all the steps of a task in one call, Jev picks the right element for each step, and wingman clicks and checks the page.
It knows its limits. When Jev isn't sure, or a step needs something wingman doesn't do (a dialog, a drag, content inside an iframe), it hands that step back to the agent instead of guessing.
Full write-up: https://hrishikesh917277.substack.com/p/handing-an-agents-browser-clicks
Repo: https://github.com/coderexpert123/jev-browser-wingman. I am not affiliated with TypeSafe.