This paper presents the feature tokeniser-transformer autoencoder (FT-TAE), a novel adaptation of the FT-Transformer’s per-feature tokenisation for inherently interpretable unsupervised tabular anomaly detection. Tailored for mixed-type tabular data, FT-TAE leverages feature-wise token reconstruction. Per-token bottlenecks produce feature-specific error signals, directly linking anomalous predictions to individual features. By reconstructing each feature token independently and maintaining feature dependencies through attention, FT-TAE promotes interpretable anomaly detection while achieving competitive detection performance and statistical gains over state-of-the-art unsupervised anomaly detectors on benchmark datasets. Model interpretability is further demonstrated on a synthetic dataset through comparative analysis with kernel SHAP, a widely adopted industry standard for feature attributions.