The measured electromagnetic field in magnetotellurics (MT) is composed of the natural source field and its subsurface response. Commonly, the data are represented as impedances, the complex ratio between the horizontal electric and magnetic fields. This measure is independent of the source distribution because the impedance-tensor estimation contains a deconvolution operator. We have used a Gauss-Newton-type 3D MT inversion scheme to compare impedance-data inversion with an inversion using the recorded electric field directly. The use of the observed electric field is beneficial to the inversion algorithm because it simplifies the estimation of the sensitivities. The direct-field approach permits the use of the observed data without processing, but it presumes knowledge of the source distribution. A method to estimate the time-variable strength and polarization of the incoming plane-wave source is presented and tested on synthetic and real-data examples. The direct-field inversion is successfully applied to a synthetic and a real data set within marine settings. A comparison with the conventional impedance inversion is conducted. The results of the synthetic data example are very similar, with a slightly more accurate reconstruction of the model in the impedance case, whereas the direct-field inversion produces a smoother inversion result when compared with the impedance case. The mapping of a resistive salt structure in the real-data example indicates deviations in the final conductivity models. The impedance inversion suggests a deeper rooted resistive structure, whereas the direct-field inversion predicts a more compact structure limited to the overburden. We have evaluated the advantages of the new approach like the simplification of the sensitivity calculation, limitations, and disadvantages like knowledge of the source distribution.