To operate effectively in unstructured environments, robots must execute language-driven, long-horizon tasks that involve both perception and interaction. This requires a 3D world representation that supports open-vocabulary understanding and manipulation-aware planning. Prior approaches often rely on closed-set detectors or focus solely on object-goal navigation, limiting their ability to handle novel object categories and manipulation contexts. To address these limitations, we propose EMO-MAN, an embodied manipulation-oriented system that integrates open-vocabulary grounding into 3D semantic mapping. EMO-MAN constructs and maintains a dynamic object-centric map with both geometric and semantic information, and selects navigation goals tailored for manipulation using both map data and task intent. This contributes to improved multi-step planning performance in dynamic and realistic environments. Code is available at https://github.com/yutian929/EMO-MAN.git