Blockchain, a decentralized and distributed ledger technology, provides stable data exchange and reliable authentication for Internet of Things (IoT) devices, introducing new ideas for the development of the IoT. Practical byzantine fault tolerance (PBFT) is well suited for IoT environments because of its low computational power requirements, but it also suffers from limitations such as poor scalability and high communication overhead. In this paper, an improved PBFT algorithm named CRPBFT, which is based on clustering and reputation values, is proposed. First, the optimized k-medoids are employed to cluster the nodes to reduce the impact of noise and improve the clustering accuracy by selecting the nodes that are farther apart as the initial clustering nodes. Second, the reputation evaluation mechanism is constructed to determine the reputation value on the basis of the performance of nodes and transaction accuracy, timeliness, cooperation activity and historical reputation stability, and nodes with high reputation are chosen. After that, the primary node is randomly selected from the high-reputation candidate nodes to reduce the possibility of the primary node being evil. Second, the consensus structure is improved by replacing nodes communicating with each other with consensus nodes communicating with primary nodes. Finally, experimental analysis compares CRPBFT and PBFT with respect to the consensus delay, communication overhead and throughput. The results show that CRPBFT has a significant advantage in terms of communication overhead, consensus delay and throughput, providing a more efficient and secure consensus scheme for blockchain applications in the IoT.