My two cents...
Not 100% accurate, but can give you a rough idea...
Linear regression makes a linearization of a problem where $y = f(x)$, with $x$ and $y$ are continuous variables.
Now imagine that you want to predict a kind of boolean behavior (yes/no) based on a $x$ value. For example, based on your salary, are you happy or not.
You can say happy = 1 and not happy = 0. You can make a scatter plot with all pairs (salary, happy) (happy in the vertical axis).
You can try to make a line to separate the happy people from the unhappy ones, but you'll see quickly that's not working well (what is a value in the middle, etc.).
A better idea would be to draw a kind of s curve which will pass as best through the points you have.
That's what the logistics regression makes. It basically makes linear the s curve by transforming the $y$ values: this is the Logit function.
We then say that we predict "True" if the predicted logit is higher than a threshold. This threshold corresponds often to 0.5, which is the inflexion point of the curve (some tools are only allowing to use 0.5). That threshold is in fact the probability.
When we have more than yes/no possibilities, one solution is to make the logistics regression for all the possibilities (e.g. if you have A, B, C, that will be A/not A, B/not B, C/not C) and take the possibility that give you highest probability. This is called the "one-vs-all" approach.