The paper introduces CodeFilling, a novel communication-efficient collaborative 3D detection system designed to enhance the perceptual abilities of agents through the exchange of perceptual messages. The system addresses the fundamental trade-off between perception ability and communication cost in collaborative perception by optimizing message representation and selection. Key contributions include: 1. **Codebook-Based Message Representation**: This method enables the transmission of integer codes instead of high-dimensional feature maps, significantly reducing communication overhead. 2. **Information-Filling-Driven Message Selection**: This approach optimizes local messages to collectively meet each agent's information demands, preventing information overflow among multiple agents. CodeFilling is evaluated on both real-world and simulation datasets (DAIR-V2X and OPV2VH+) in homogeneous and heterogeneous settings. Results show that CodeFilling outperforms previous state-of-the-art methods (Where2comm) with 1.33 and 1.206 times less communication volume, respectively. The system maintains superior performance in both homogeneous and heterogeneous settings, demonstrating its robustness and efficiency in collaborative perception tasks.The paper introduces CodeFilling, a novel communication-efficient collaborative 3D detection system designed to enhance the perceptual abilities of agents through the exchange of perceptual messages. The system addresses the fundamental trade-off between perception ability and communication cost in collaborative perception by optimizing message representation and selection. Key contributions include: 1. **Codebook-Based Message Representation**: This method enables the transmission of integer codes instead of high-dimensional feature maps, significantly reducing communication overhead. 2. **Information-Filling-Driven Message Selection**: This approach optimizes local messages to collectively meet each agent's information demands, preventing information overflow among multiple agents. CodeFilling is evaluated on both real-world and simulation datasets (DAIR-V2X and OPV2VH+) in homogeneous and heterogeneous settings. Results show that CodeFilling outperforms previous state-of-the-art methods (Where2comm) with 1.33 and 1.206 times less communication volume, respectively. The system maintains superior performance in both homogeneous and heterogeneous settings, demonstrating its robustness and efficiency in collaborative perception tasks.