Code-mixed language is highly prevalent in South Asian social media and poses challenges for automated toxicity detection due to intra-sentential language mixing, transliteration noise, and orthographic variation. This study proposes MCT-Mix, a Multilingual Transformer-BiLSTM hybrid model designed to detect toxicity in mixed-coded posts spanning four language subsets: Bengali (native script), Hindi (native script), Banglish (Bengali in Latin script), and English. These subsets are internally mapped to a three-tag language identification scheme, Bengali (bn), English (en), and other (oth), where oth captures Hindi and Banglish tokens for consistent feature extraction. The architecture employs four parallel transformer encoders to generate contextual embeddings, followed by a Bidirectional Long Short-Term Memory (BiLSTM) layer to model cross-lingual sequential dependencies. To support evaluation, we introduce MixTox-SA, a balanced corpus of 1 3, 0 3 5 manually annotated Facebook comments labeled as Toxic, Non-Toxic, or Neutral. Experiments against Random Forest, vanilla BiLSTM, and MuRIL baselines show that MCT-Mix achieves a macro-F1 of 91.53 and AUC above 0.97, outperforming standalone MuRIL by 9.1 points. These results demonstrate the model's effectiveness in capturing semantic meaning in noisy code-mixed text for multilingual toxicity detection.