This paper introduces HEterogeneous ALliance (HEAL), a novel extensible framework for open heterogeneous collaborative perception. The framework addresses the challenge of integrating new heterogeneous agent types into existing collaborative systems while ensuring high perception performance and low integration costs. HEAL establishes a unified feature space using a multi-scale foreground-aware Pyramid Fusion network and aligns new agents to this space through a backward alignment mechanism. This approach minimizes training costs and preserves model details, making it suitable for real-world applications. Extensive experiments on the OPV2V-H and DAIR-V2X datasets demonstrate that HEAL outperforms state-of-the-art methods in collaborative detection performance while reducing training parameters by 91.5% when integrating three new agent types. The framework's effectiveness is further validated through robustness experiments against pose errors and feature compression.This paper introduces HEterogeneous ALliance (HEAL), a novel extensible framework for open heterogeneous collaborative perception. The framework addresses the challenge of integrating new heterogeneous agent types into existing collaborative systems while ensuring high perception performance and low integration costs. HEAL establishes a unified feature space using a multi-scale foreground-aware Pyramid Fusion network and aligns new agents to this space through a backward alignment mechanism. This approach minimizes training costs and preserves model details, making it suitable for real-world applications. Extensive experiments on the OPV2V-H and DAIR-V2X datasets demonstrate that HEAL outperforms state-of-the-art methods in collaborative detection performance while reducing training parameters by 91.5% when integrating three new agent types. The framework's effectiveness is further validated through robustness experiments against pose errors and feature compression.