The Gaia mission has provided the largest ever astrometric chart of the Milky Way. Using it to map the Galactic halo is helpful for disentangling its merger history. The identification of halo stars in Gaia DR2 with reliable distance estimates requires special methods because such stars are typically farther away and scarce. We applied the reduced proper motion (RPM) method to identify halo main sequence stars on the basis of Gaia photometry and proper motions. Using the colour-absolute-magnitude relation for this type of star, we calculated photometric distances. Our selection results in a set of 10^7^ tentative main sequence halo stars with typical distance uncertainties of 7% and with median velocity errors of 20km/s. The median distance of our sample is ~4.4kpc, with the faintest stars located at ~16kpc. The spatial distribution of the stars in our sample is centrally concentrated. A visual inspection of the mean velocities of stars on the sky reveals large-scale patterns as well as clear imprints of the GD-1 stream and tentative hints of the Jhelum and Leiptr streams. Incompleteness and selection effects limit our ability to interpret the patterns reliably as well as to identify new substructures. We define a pseudo-velocity space by setting the line-of-sight velocities of our sample stars to zero. In this space, we recover several known structures such as the footprint of Gaia-Enceladus (i.e. the Gaia-Sausage) as well as the Helmi Streams and some other retrograde substructures (Sequoia, Thamnos). We show that the two-point velocity correlation function reveals significant clustering on scales smaller than 100km/s of a similar amplitude as found for the 6D Gaia halo sample. This clumping of stars in velocity space might hint at the presence of nearby streams that are predominantly phase-mixed. A spectroscopic follow-up of our halo main sequence sample is bound to yield unprecedented views of Galactic history and dynamics. In future Gaia data releases, the level of systematics will be reduced and the astrometry will be more precise, which will allow for the identification of more substructures at larger distances.