This paper describes a server selection mechanism for connection oriented services based on passive probing. The criterion of selection is the quality of service expected from each server, expressed as a function of availability and response time. Measures from previous connections to servers made by local clients are used to continuously update a QoS database which the prediction algorithm uses to compute the response time expected in subsequent connections. The forecasting approach is mainly based on prior measurements of TCP connection establishment time. The maximum segment size in a connection is also considered. The proposed metric is compared with other ones normally used to measure network proximity. Results show that the proposed server selection mechanism achieves a reduction of response time of over 50 percent compared with a random selection mechanism.