I am not familiar with Reiter’s method, but I do not understand what you are looking for here. Dynare obtains the first order approximation by computing the Jacobian of the model symbolically. What would be the advantage of using automatic differentiation here?
If you have a code returning the residuals and jacobian of the dynamic model, you could in principle replace the matlab routines created by Dynare (using symbolic calculus) provided that you manage to use the same API (and that the Jacobian matrix is organized as done in Dynare). But I fail to see the point. Unless you aim to pass the automatic differentiation objects to Dynare resolution routines… But this will most likely not work (unless you decide to not use the mex routines used for computing the reduced form solution, which clearly cannot eat these objects, and that your automatic differentiation class manage to work with simple algebra operations like matrix multiplication and inversion).