BACKGROUND: Workplace-based mindfulness programs have good evidence for improving employee stress and mental health outcomes, but less is known about their effects on productivity and citizenship behaviors. Most of the available evidence is derived from studies of mindfulness programs that use class-based approaches. Mindfulness apps can increase access to training, but whether self-directed app use is sufficient to realize benefits equivalent to class-based mindfulness programs is unknown. OBJECTIVE: We assessed the effectiveness of a mindfulness app, both with and without supporting classes, for reducing employees' perceived stress. Changes in mindfulness, mental health, quality of life, perceptions of job demand, control and support, productivity indicators, organizational citizenship, and mindful behaviors at work were also investigated. METHODS: Tasmanian State Service employees were invited by the Tasmanian Training Consortium to a 3-arm randomized controlled trial investigating the effects of a mindfulness app on stress. The app used in the Smiling Mind Workplace Program formed the basis of the intervention. The app includes lessons, activities, and guided meditations, and is supported by 4 instructional emails delivered over 8 weeks. Engagement with the app for 10-20 minutes, 5 days a week, was recommended. Reported data were collected at baseline (time point 0), 3 months from baseline (time point 1 [T1]), and at 6-month follow-up (time point 2). At time point 0, participants could nominate a work-based observer to answer surveys about participants' behaviors. Eligible participants (n=211) were randomly assigned to self-guided app use plus four 1-hour classes (app+classes: 70/211, 33.2%), self-guided app use (app-only: 71/211, 33.6%), or waitlist control (WLC; 70/211, 33.2%). Linear mixed effects models were used to assess changes in the active groups compared with the WLC at T1 and for a head-to-head comparison of the app+classes and app-only groups at follow-up. RESULTS: App use time was considerably lower than recommended (app+classes: 120/343 minutes; app-only: 45/343 minutes). Compared with the WLC at T1, no significant change in perceived stress was observed in either active group. However, the app+classes group reported lower psychological distress (β=-1.77, SE 0.75; P=.02; Cohen d=-0.21) and higher mindfulness (β=.31, SE 0.12; P=.01; Cohen d=0.19). These effects were retained in the app+classes group at 6 months. No significant changes were observed for the app-only group or for other outcomes. There were no significant changes in observer measures at T1, but by time point 2, the app+classes participants were more noticeably mindful and altruistic at work than app-only participants. CONCLUSIONS: Including classes in the training protocol appears to have motivated engagement and led to benefits, whereas self-guided app use did not realize any significant results. Effect sizes were smaller and less consistent than meta-estimates for class-based mindfulness training. TRIAL REGISTRATION: Australian New Zealand Clinical Trials Register ACTRN12617001386325; https://www.anzctr.org.au/Trial/Registration/TrialReview.aspx?id=372942&isReview.