The goal of few-shot classification (FSC) is to use only a small number of labeled data points (typically 1 to 5) per class to train a classification model. It is not always possible to obtain a large number of labeled samples for many real-world scenarios while plenty of unlabeled data is available. In this work, we propose MetaFix, a semi-supervised model agnostic meta-learning framework designed to boost the performance of semi-supervised FSC. To leverage unlabeled samples, the proposed method incorporates consistency regularization and uses pseudo-labeling to enhance the performance of semi-supervised FSC. Our approach also effectively utilizes unlabeled data both in the outer and inner loops of MAML algorithm, resulting in more robust and informative meta-learned model parameters. Additionally, it effectively mitigates the effects of misclassification and noise-influenced data since it employs consistency regularization, which preserves semantics. We study the performance of MetaFix for the case where the unlabeled dataset belongs to the same dataset as the labeled data in each episode. We evaluate our approach on tieredImageNet, miniImageNet, and CIFAR-FS datasets for various settings. Our experiments demonstrates that the proposed approach achieves better performance than MAML and other semi-supervised few-shot learning methods.