GAMEHARBOR: Design and Development of a Casual Game Platform with an AI Wellbeing Companion for Everyday Mental Health

Background. Nearly a billion people were living with a mental disorder in 2019, and most never receive care. Mobile mental health apps could help close that gap, but people rarely stay with them: the median app keeps fewer than 4% of its users after 15 days. Casual games, by contrast, are things people return to by choice, and there is modest evidence that playing them can lift mood and ease stress. Conversational agents have also begun to show measurable effects on depressive symptoms and distress. Very few tools bring the two together, and fewer still are built with safety at the centre. Objective. This paper describes the design and development of GAMEHARBOR, a free app for Android, iOS and the web that pairs a library of short, calming games with Ollie, an animated octopus companion who listens, guides evidence-based coping skills and points people to real help when they need it. It sets out the evidence behind each design decision, describes how the system works, reports the results of development-stage testing, and proposes a study to measure its effect on wellbeing. Methods. Design decisions were drawn from published research on games and wellbeing, serious games, psychotherapy, conversational agents and app engagement, and from NHS and World Health Organization guidance. The app was built as a single JavaScript code base packaged with Capacitor for mobile and as a progressive web app. Ollie combines an on-device conversation engine with optional replies from an open large language model (Google Gemma) running on the app's own server, grounded in a curated library of 120 short reference notes. A layered safety system detects crisis language on the device, asks directly about suicidal thoughts when someone sounds hopeless, gives country-specific helplines, and stops suggesting games when someone is struggling. Development testing combined scripted on-device tests with automated evaluation of live AI replies against fixed rules. Results. Version 1.12.10 (October 2026) includes 62 games across five categories, guided breathing with an optional microphone breath check, mood and journal tools, and two standard self-check questionnaires (PHQ-9 and GAD-7) delivered inside the conversation. In the most recent automated evaluation rounds, 21 of 22 live AI replies met every rule; the one failure led to a fix that was confirmed on retest. All scripted safety, check-in and turn-taking tests passed after fixes. Conclusions. GAMEHARBOR shows that a games-first wellbeing app can be built with clinical safety principles, open AI models and privacy by default, at very low running cost. Whether it improves wellbeing is not yet known; a pilot study is proposed to answer that question. This is a preprint and has not been peer reviewed. GAMEHARBOR is a wellbeing tool, not a medical treatment.