I’ve been trying headroom-ai for this reason. That project also reduces tokens in other clever ways. Have you looked into it?