Service recommendation for mashup development faces critical challenges due to the sparse usage history of newly introduced mashups and APIs, as well as the difficulty of inferring genuine API dependencies from mashup compositions, where APIs are often co-used in a noisy and implicit manner. Traditional collaborative filtering, content-based methods, and standalone LLM-based approaches have limitations in jointly addressing these challenges within a unified framework. We propose MARS, a multi-agent collaborative recommendation framework that systematically integrates semantic alignment, structure-aware retrieval, and validation-based recommendation under a constrained candidate space. MARS incorporates multiple algorithmic components to improve different stages of the service recommendation process. Specifically, agent-driven semantic enrichment substantially mitigates cross-representation semantic mismatch between mashups and APIs, reducing the average Jensen-Shannon distance from 0.7333 to 0.6333, while baseline methods exhibit negligible changes. Structure-aware fine-tuning captures API compositional patterns beyond surface-level semantics, and data-driven weight optimization learns the fusion weights in the hybrid retrieval stage, replacing static retrieval parameters with empirically calibrated strategies. Finally, multi-agent collaborative reasoning enhances robustness by combining diverse proposal generation with validation-based selection. Experiments on ProgrammableWeb dataset demonstrate that MARS consistently outperforms representative baselines, achieving 63.31% Recall@5 compared to 58.28% for Native RAG and 43.35% for the best traditional method (ServeNet). The results indicate that MARS provides an effective and extensible framework for improving mashup-oriented service recommendation. Our code is available athttps://github.com/banirabbit/mars.