Probability distribution as output of a random forest model

Question

Is this possible?

I am going to use Tensorflow decision trees, but if any other implementation has this feature I would switch

To make it easier, I would want to use logits as output, and then softmax them into a probability distribution function.

Going further, I want to do have 'multiple output heads'. One head would be the aforementioned pdf, but another would just be a single regression variable 'value' (think alpha zero.)

I can imagine how this would work algorithmically. The leafs of the decision tree would simply have a vector (logits, value). The logits for each leaf can be converted into a pdf, and the average taken as the output. Same essentially goes for the value.

Any advice or references are welcome.