Cokriging is a form of spatial interpolation and requires that you have a coordinate for every data point you want to include. A really good paper is "Cokriging model for estimation of water table elevation" (Hoeksema et al., 1989) - just the introduction can help to get a feel for what the algorithm is doing. Cokriging also requires that at least one data point from each data set be at the same location, or at least very close.
If you have a categorical variable, you need to define it spatially. If you feel that the variable is more or less equally weighted across your study area, you could set up a grid of data points. To do this, you could create a fishnet, then use the extract values to points to pull the data values from your raster file into your point grid. If you have more detailed information about the spatial distribution of your categorical variable, you may want to manually select points or looking into methods to create a weighted grid. If you don't have a raster of your categorical data but you do have sample locations, that would be perfect to include, and no grid creation is required. If your categorical variable is of a non numerical data type, then use a numerical code for each individual type in your data set (Ex: "aster" = 1). Better yet, use a measured or known property from each vegetation type, like average leaf area index, rooting depth, etc.
If you want to use cokriging to interpolate the vegetative community, use this as your primary variable and your continuous data as a secondary input. Your continuous variable should also be defined as a grid of data points, which you can do using the fishnet method I described above.
A really great book for understanding spatial statistics is "An Introduction to Applied Geostatistics" by Isaaks and Srivastava. You will find this text referenced in many spatial statistics studies.