Virtual clients provide a cost‑effective way to support A/B testing, recommender development and interface evaluation. Building them, however, requires large‑scale, semantically faithful, fine‑grained online user trajectories, which are hard to obtain due to privacy constraints and limited traffic for many businesses. Existing public datasets either abstract away interaction details or preserve rich context only for specific platforms and at small scale.\ \ To bridge this gap, we introduce SimTrace, a framework that generates fine‑grained synthetic multimodal clickstreams using a computer‑use client agent grounded in real user trajectories and the target web environment. SimTrace first anonymizes real interactions and constructs a simulated twin of the web environment; the agent then operates within this twin, pairing each action with the corresponding web observation and user context. The result is a shareable synthetic log that can replace confidential real logs for developing computer‑use‑style virtual clients.\ \ We apply SimTrace to an e‑commerce setting and evaluate it on eight fidelity metrics against several baselines, achieving superior performance on seven of them. Models trained on the synthetic data attain comparable results to those trained on real data for downstream tasks such as purchase prediction and recommendation. For next‑action prediction, augmenting real data with synthetic data improves accuracy by 11.0% relative to using real data alone.\ \ SimTrace is released as an open‑source package to facilitate research on online user behavior modeling.\ \ Review