The introduction of direction finding in the Bluetooth standard enabled the use of antenna arrays for locating Bluetooth devices, using carrier phase measurements to estimate the direction from the array to a moving device. In this work, this feature is utilized for outdoor localization. We show how using repeated measurements from all array elements, instead of only the initial single-element reference samples as often suggested, can contribute to an improved estimate of the signal’s unknown carrier frequency offset, thereby improving the direction estimation performance. To run the direction-of-arrival estimation in real-time with high angular resolution on an embedded computer we propose a pseudo-spectrum peak search strategy that combines a coarse search, where the resolution is decided based on the array’s main lobe width, with a local nonlinear optimization for estimate refinement. We consider practical aspects relating to the phase sampling configuration and demonstrate direction estimation at up to 700m range with insignificant packet loss within 500m, and without significant loss of angular precision even when the received signal is near the receiver’s signal strength sensitivity threshold. In an open outdoor environment, using a square antenna array with 12 elements, the azimuthal performance is found to be very consistent with range, with noise standard deviation typically around 1°. While the elevation is significantly affected by multipath at lower elevation angles, with visible disagreement between frequency channels, it is shown to be consistent with simulations of ground reflection multipath.