Artificial intelligence (AI), particularly through large language models (LLMs), is transforming engineering by making advanced tasks - that previously required specialized expertise - accessible to a wider audience. In line with this shift, we aim to enhance the accessibility of multibody dynamics (MBD) using virtual labs. Our primary objective is to extend open MBD knowledge by generating a comprehensive knowledge base using LLMs. This knowledge base shall be represented by datasets containing multibody models, their simulation codes and corresponding validated conjectures. In this context, the model precisely defines the general structure and components of the multibody system along with the parameters. Additionally, the dataset includes information about the corresponding code required to simulate the model, such as Python code for Exudyn models. Conjectures (e.g., the slider moves along a line) are repeatedly generated for each model - either randomly or using reliable engineering sources - and validated (or disproved) utilizing the simulation model. In a simplified variant, we generate generic multibody models of various kinds, which shall serve as examples and can be used with in-context learning approaches. The present approach is enabled by the capabilities of state-of-the-art open and closed source LLMs, which already demonstrated that they can generate simulation code from natural language. As a general approach, we introduce AI agents within virtual laboratories. These agents aim to iteratively perform computational virtual experiments to validate or disprove LLM-generated conjectures and store results within a database. Similar to real-world experiments, initially stated conjectures are accepted or rejected with the help of simulation outputs. A dataset is added to the knowledge base if it contains a large number of validated conjectures, as it is assumed to contain both a correct conjecture as well as correct simulation models. First tests have shown that base models like Meta's 70B-parameter Llama 3.1 instruct model are capable of creating simulation code relevant to accepting or rejecting a conjecture. Upon the generation of larger datasets, virtual engineering labs may represent a paradigm shift in engineering practice and outcomes may be prolonged to other engineering areas.