Given that you would need an exceedingly extensive, labeled dataset to start working with actions/behaviours/body language directly, the only realistic way to start would be with a language model. (And while these are sometimes decent, they're absolutely terrible at other times (and there's generally no way to tell which time is which).)

Anything Linux based is out, but I feel like there might be some surviving bsd that's a descendant of one of the paid versions. That being said, GNU utils are out and I don't know of any browsers that would work either (though there are utils to download and look at the source that would work).