Abstract We investigated the plumbing system of Stromboli volcano from the upper mantle to the surface. Thermobarometric estimates indicate that the deeper detected part of the plumbing system is located in the upper mantle, at approximately 34–24 km depth where, during their ascent, primitive Stromboli basalts (HKCA to shoshonitic) interact with peridotitic materials. In this region magma flow is probably channelled along fracture zones that may converge into a feeder dyke that crosscuts the Moho at about 17 km depth. During their ascent, basaltic magmas will interact with lower crust materials represented by cumulates of earlier Stromboli-type basalts at 13–10 km depth. This zone is also the section of the plumbing system where the feeder dyke is entering the chamber. Thermobarometric estimates, obtained by constructing a grid of selected reactions, indicate that current primitive Stromboli basalts equilibrate at 0.3–0.15 GPa and temperatures approaching 1200 °C, and progressively crystallize and degas before being erupted. Crystal size distributions on lavas and juvenile tephra erupted in 2002–2003 give very variable residence times. Based on average bubble distances, the estimated times for the exsolution of the gaseous phases range from 2–7 days to 45 min for the lavas and scorias, down to about 15 h to 12 min for the pumices erupted during paroxysmal explosions. Estimated syneruptive viscosities range from 10 2 Pa s for the anhydrous basaltic pumices at 1200 °C, to 10 3 –10 4 Pa s for lavas approaching their effusion temperatures (1100–1150 °C). In turn, viscosities for the hydrous basaltic melt that led to the formation of the basaltic pumices may be around 10 Pa s or lower. In the light of the above, we discuss the possible shapes and volumes of Stromboli magma chamber by considering a sphere, an ellipsoid (geometrically concordant with the regional stress distribution) and a feeder dyke, the last two being more likely. In the light of volcanological, structural and geophysical data on conduit thickness, we propose an alternative model that takes into account the volumes of recently erupted lavas. This model consists of a convective ellipsoidal magma chamber ‘injected’ by an active feeder dike of undegassed magma of higher temperature, lower density and lower viscosity. This dyke will evolve into a magma column inside the chamber and will separate the reservoir into two lateral, nearly symmetric convective regions. Crystallization would occur preferentially in the proximity of the wallrocks, particularly where the chamber is entering the conduit. The onset of paroxysmal explosions during major effusive cycles may be explained by a drastic increase in the intrusion rates at the base of the chamber that will produce a progressive inflation of the magma column dynamically transferred to the chamber walls. The ceasing of ‘anomalous’ intrusion rates at the base of the chamber, coupled with higher discharge rates, will progressively depressurize the chamber to a critical threshold, until the stress transferred to the walls is dynamically released: at this point the walls themselves will undergo a nearly instantaneous elastic rebound and contract in the attempt to recover their original pre-eruptive geometry. These dynamics will squeeze up portions of the undegassed magma column, triggering a paroxysmal explosion with the ejection of ‘golden pumices’.