Kult: A Gemini-powered gaming companion that knows your game

Kult is a London-based startup building a companion that sits alongside a gaming session. It understands which game a player is in and where, answers by voice, and shows a visual feed on a second screen. The stated problem is that mid-game help traditionally means alt-tabbing, videos, forums, and losing context. Kult wants information to come to the player, with hands on keyboard or controller, and without breaking immersion.

Three constraints shaped the design: latency (answers must be fast enough to feel like an expert friend next to you), accuracy (general-purpose models have training cutoffs, so they can be wrong about a specific game, patch, or build; answers must also depend on player progress and play style), and cost (an always-on companion handles many turns per session, and covering hundreds of games requires adding one to be cheap and automated). The team chose Gemini models as the default family on both sides of the system: live conversation and knowledge reasoning. The lightest Gemini Flash tier handles high-volume repetitive work such as conversational turns and grounding answers in Kult‘s knowledge. The most advanced Gemini models handle rare heavy reasoning done once per game. Models are reached through the Gemini platform and Gemini Developer API, with a routing layer and a backend on Google Kubernetes Engine.

Every answer has two parts. The voice gives a short version; at the same moment, a visual answer lands in a feed on the player’s second screen with images, embedded video, maps, builds, interactive tools, and links to guides and creators. The feed is built with Google’s open standard for agent-generated interfaces: the model does not write code but describes the interface using a catalog of components, and the app renders it natively in Kult‘s style. That keeps the feed safe, consistent, and extensible.

Kult now covers more than 100 games, and adding a game is an automated process, not a manual project. It is in alpha with first players and opening up to more. The article describes five principles. First, answers come from a structured knowledge base built automatically for each game and kept current with patches, rather than from what a model happens to remember. Second, the answer depends on who is asking: Kult tracks the session, including the game, recent actions, and prior questions, so references like “What is he weak to?” work without clarification. It learns from play style over time. Third, the player is never left in silence: the companion responds right away and carries the conversation while deeper reasoning happens behind it; the spoken answer is short and the visual feed carries what voice cannot. Fourth, the smallest model that can do the job does the job. Many tasks are decisions, not writing. For those, Kult uses Jev, a decision-only model that returns a choice with a calibrated confidence score; confident cases settle in a fraction of a second at lower cost, and uncertain ones escalate to a larger Gemini model. Every model sits behind a registry, so new generations are benchmarked against reference answers written by people who know the games before being swapped in. Fifth, every question makes Kult better: player questions reveal where a game’s knowledge is thin, so the team builds what to add next. The authors think the companion that sits beside the game is only the start.

Next steps: computer vision. In coming months, Kult will start to see what is happening in the player’s game, so the companion can be aware of the moment without being asked. The team is using Google’s Live models for video understanding and more. The article also notes that Kult works closely with the Google Cloud team and has direct access to engineers, which matters when every model generation raises what a gaming companion can do.

Kult built a gaming companion with Gemini that actually knows your game

View Original