We have de ne a network programming model called Mobile Assistant Programming (MAP) for development and execution of communication applications in large scale networks of heterogeneous computers. MAP assistants are high-level interpreted programsthat can movebetween nodes, create clones and report results. Their execution is asynchronous and persistent to take into account client disconnections and node failures. This paper presents the implementation of the MAP model using the World-Wide Web framework and the Scheme programming language. Our measures and the rst experience with MAP applications show that signi cant performance improvement can be achieved by moving computation closer to data: both the elapsed time and network tra c are reduced.