A client-side dispatcher resides on a client machine below high-level client applications and TCP/IP layers. The client-side dispatcher performs TCP state migration to relocate the client-server TCP connection to a new server by storing packets locally and later altering them before transmission. The client-side dispatcher operates in several modes. In an error-recovery mode, when a server fails, error packets from the server are intercepted by the client-side dispatcher. Stored connection packet's destination addresses are changed to an address of a relocated server. The altered packets then establish a connection with the relocated server. Source addresses of packets from the server are changed to that of the original server that crashed so that the client application is not aware of the error. In a delayed URL-based dispatch mode, the client-side dispatcher intercepts connection packets before they are sent over the network. Reply packets are faked by the client-side dispatcher to appear to be from a server and then sent to up to the client TCP/IP layers. The client's TCP then sends URL packet identifying the resource requested. The client-side dispatcher decodes the URL and picks a server and sends the packet to the server. Reply packets from the server are intercepted, and data packets altered to have the source address of the faked server. Multicast of the initial packet to multiple servers is used for empirical load-balancing by the client. The first server to respond is chosen while the others are reset. Thus the client-side dispatcher picks the fastest of several servers.