The hyperdynamics method is a powerful tool to simulate slow processes at the atomic level. However, the construction of an optimal hyperdynamics potential is a task that is far from trivial. Here, we propose a generally applicable implementation of the hyperdynamics algorithm, borrowing two concepts from metadynamics. First, the use of a collective variable (CV) to represent the accelerated dynamics gives the method a very large flexibility and simplicity. Second, a metadynamics procedure can be used to construct a suitable history-dependent bias potential on-the-fly, effectively turning the algorithm into a self-learning accelerated molecular dynamics method. This collective variable-driven hyperdynamics (CVHD) method has a modular design: both the local system properties on which the bias is based, as well as the characteristics of the biasing method itself, can be chosen to match the needs of the considered system. As a result, system-specific details are abstracted from the biasing algorithm itself, making it extremely versatile and transparent. The method is tested on three model systems: diffusion on the Cu(001) surface and nickel-catalyzed methane decomposition, as examples of “reactive” processes with a bond-length-based CV, and the folding of a long polymer-like chain, using a set of dihedral angles as a CV. Boost factors up to 109, corresponding to a time scale of seconds, could be obtained while still accurately reproducing correct dynamics.