Many animal behaviours involve complex three-dimensional movements, yet detailed 3D analyses have remained rare, because available tools are tied to a few body plans - typically humans or a handful of mammals - and often require substantial programming expertise. To address this gap, we developed SMILify, an open-source framework extending parametric body modelling to virtually any rigged 3D mesh. A custom Blender Add-on enables researchers to build parametric animal models from their own 3D scan collections. From these, the system automatically learns what range of body proportions, limb lengths, and overall morphologies are anatomically plausible for a given (set of) species, and constrains reconstructions to remain within these bounds. SMILify leverages this learned shape space within a neural inference pipeline that takes either 2D pose estimates from existing tracking tools (e.g. SLEAP or DeepLabCut) or triangulated 3D key-points from calibrated multi-camera arrays to produce dense, dynamic 3D reconstructions of body pose and shape during behaviour, learning first from synthetic data, then adapting to real recordings using the same pose estimates researchers have already generated. Applications in stick insects, ants, and mice, across single- and multi-camera systems, demonstrate the framework's potential for morphometric studies of shape variation, kinematic decomposition of gait, and long-duration behavioural monitoring without physical markers. By enabling researchers to build species-specific models from their own data to then reconstruct 3D pose and shape from existing video, we hope to lower the barrier to quantitative 3D analysis of animal behaviour in the lab, in the field, and across taxa.