Continuous monitoring of the position of a subject's head is an essential part of improving localization accuracy and resolution in MEG. We describe a procedure that has been developed for whole-cortex MEG sensors. The system uses three (or more) small head coils driven continuously by low-amplitude sinusoidal currents with frequencies chosen so they do not interfere with MEG measurements and with each other and are easily separated from power-line signals and harmonics. Analysis of the response of the MEG sensors to the head coils allows continuous monitoring of the position (update times as short as T=2/fpower) using a 3-parameter minimization. The best-fit positions of the head coils are then combined to determine the head translation and rotation. Analysis of phantom data recorded with a 275-channel CTF MEG system in a shielded room shows that coil positions can be determined with an accuracy of approximately 2 mm with an update period T=1/15 s even when the head coils are moving approximately 25 mm at speeds up to 40 mm/s. Data are corrected by expressing the scalar potential for the magnetic field as a spherical-harmonic series, and then determining the effect of rotations and translations on the terms of the series. Since the MEG helmet covers only approximately 60% of the full sphere, care must be taken in determining the coefficients of the spherical-harmonic series to ensure that the modeled magnetic field does not become unrealistically large in the region where there are no MEG sensors (i.e. in the lower 40% of the sphere). Our approach has been to use a minimum-field-energy criterion that minimizes the squared gradient averaged over 4pi sr and radii from 96 to 145 mm while matching the MEG measurements.