Web-based education and training provides a new paradigm for imparting knowledge; students can access the learning material anytime by operating remotely from any location. Web3D open standards, such as X3D and VRML, support Web-based delivery of Educational Virtual Environments (EVEs). EVEs have a great potential for learning and training purposes, by allowing one to circumvent physical, safety, and cost constraints. Unfortunately, EVEs often leave to the user the onus of taking the initiative both in exploring the environment and interacting with its parts. A possible solution to this problem is the exploitation of virtual humans acting as informal coaches or more formal instructors. For example, virtual humans can be employed to show and explain maintenance procedures, allowing learners to receive more practical explanations which are easier to understand. However, virtual humans are rarely used in Web3D EVEs, since the programming effort to develop and re-use them in different environments can be considerable. In this paper, we present a general architecture that allows content creators to easily integrate virtual humans into Web3D EVEs. To test the generality of our solution, we present two practical examples showing how the proposed architecture has been used in different educational contexts.