This is a review paper addressing two main aspects of DNA computing research: DNA computing in vitro (in the test tube) and in vivo (in a living organism). We describe the first successful in vitro DNA computing experiment [L.M. Adleman, Science 266 (1994) 1021–1024] which solved a mathematical problem, the Directed Hamiltonian Path Problem, solely by manipulation of DNA strands in test tubes. We then address DNA computing in vivo by presenting a model proposed by Head [in: G. Rozenberg, A. Salomaa (Eds.), Lindenmayer Systems, Springer, Berlin, 1991, pp. 371–383] and also by Landweber and Kari [in: L. Kari, H. Rubin, D.H. Wood (Eds.), Biosystems, Vol. 52, Nos. 1–3, Elsevier, Amsterdam, 1999, pp. 3–13] and developed by Landweber and Kari [in: L.F. Landweber, E. Winfree (Eds.), Evolution as Computation, Springer, Berlin, 1999], for the homologous recombinations that take place during gene rearrangement in ciliates. Results given by Kari, Kari and Landweber [in: J. Karhumaki, H. Maurer, G. Paun, G. Rozenberg (Eds.), Jewels are Forever, Springer, Berlin, 1999, pp. 353–363] and Landweber and Kari [in: L.F. Landweber, E. Winfree (Eds.), Evolution as Computation, Springer, Berlin, 1999] have shown that a generalization of this model that assumes context-controlled recombinations has universal computational power. We review results obtained by Kari and Kari [in: Words, Sequences, Languages: Where Computer Science, Biology and Linguistics Meet, Kluwer Academic Publishers, The Netherlands, in press] on properties of context-free recombinations and characterize the languages generated by context-free recombination systems. As a corollary, we show [J. Kari, L. Kari, in: Words, Sequences, Languages: Where Computer Science, Biology and Linguistics Meet, Kluwer Academic Publishers, The Netherlands, in press], that context-free recombinations are computationally weak, being able to generate only regular languages.