This paper presents OpenRehabAgent, a modular multi-agent prototype for video-based pain localisation and adaptive exercise recommendation. The system links pose estimation, simple pain heuristics, reinforcement-learning-based exercise selection, user feedback, and a rule-based safety supervisor through a shared knowledge base. The current implementation runs on synthetic pose sequences and simulated users only. It is not a medical device and is not intended for clinical decision-making. The paper describes the architecture, the selection loop, and an evaluation plan using synthetic users, and outlines limitations and directions for future work.