Local seismic tomography is a well-known and commonly used method for obtaining detailed information about the internal structure of volcanoes. The eruption of Mt. Agung in 2017 was a vital opportunity scientifically because it is the first eruption that had sufficient seismic observation networks to carry out local seismic tomography at this volcano. In this study, we investigate the subsurface structure of Mt. Agung in Bali, which is one of the highest risk volcanoes in Indonesia. We conducted travel-time tomography using P- and S-wave arrival times of volcano-tectonic (VT) events to determine the three-dimensional (3D) Vp, Vs, and Vp/Vs ratio structure beneath Mt. Agung. We used 1,926 VT events, with corresponding 9482-P and 8683-S wave arrival times recorded by eight seismic stations over an observation time spanning from October 18 to December 31, 2017. We obtain the hypocenter solution for VT events using the maximum likelihood estimation algorithm and use an optimum 1D velocity model as input for the Joint 3-D seismic tomographic inversion. Local earthquake tomography revealed five anomalous regions that are useful to describe the overall seismic activity around Mt. Agung. We interpret these anomalous regions qualitatively due to limited data resolution in this study. We have successfully localized a high Vp/Vs ratio (∼1.82), low Vs (−1.9%) and high Vp (+3.8%), within a low seismicity zone at depths between 2 and 5 km below the Mt. Agung summit, which may be related to a shallow magma reservoir. There is also an anomalous region between Mt. Agung and Batur with moderate to high Vp/Vs ratios (1.76–1.79) where most of the earthquakes recorded before the 2017 eruption originated. We interpret this anomaly to be related to the existence of sub-vertical dyke complex at depths between 8 and 14 km. The results of our study provide new insights into the subsurface structure of the magma plumbing system beneath Mt. Agung, which can be used to improve the quality of determining the location of the hypocenter and source modeling for future eruption forecasting.