If you naively apply the algorithm you need an encoding $f$ of strings of length $m$ into integers such that $s_1≤s_2$ iff $f(s_1)≤f(s_2)$. The bound on these integers is necessarily at least $c^m$ where $c$ is the number of possible characters.
You then have a complexity in $O(nc^m)$ where $n$ is the number of elements to be sorted. This could be interesting when $c^m$ is not too big compared to the usual bound $\log n$, which is possible, in very peculiar cases.
In general, one would prefer the bound $O(n\log n)$ which does not depend on the length of your strings and in general much faster than $O(nc^m)$. To give a formal comparison between them, the former would be $O(s\log s)$ and the latter $O(sc^s)$ where $s$ is the size of the input.
However what is possible if you know that you will have a number of different strings no greater than some $k$, and that there will be a lot of occurrences of the same strings, is to build a correspondence between these strings and $\{1,\dots,k\}$ on which you will apply the counting sort, thus obtaining a complexity of $O(kn)$ i.e. linear (since $k$ is constant). But in this case the counting sort is not the only one reaching a linear complexity, for example all algorithms using balanced trees can be adapted to be linear, even insertion sort which is quadratic in the worst case, becomes linear.