Magicweave

Djinn turns one sentence into a real game, live online, with its own economy. Making a game today takes a team weeks: design, art, code, a backend for coins and leaderboards, and hosting. Most ideas never get past the first step. AI code tools help with pieces, but they hand you files, not a finished game, and they don't know how a game's economy should work. With Djinn, you type a wish ("a lighthouse keeper guiding ghost ships home before the lamp runs dry") and are playing a starter within seconds. Djinn then asks only the questions that change the build most, learns your taste from a short swipe through looks, and builds two 30-second versions so you can pick the way it plays before any real time is spent. A team of AI agents builds the full game in parallel: a lead writes the plan, and specialists for mechanics, levels, art and economy each own their part. Before anything ships, every promise in the plan is actually played and checked, the look is measured, and a separate judge model rates the taste. The game's coins, stats and leaderboards are set up on Magicweave's backend, so the server, not the game, decides what a run is worth. Going live is one click, and you keep refining by chatting: "make it bold and chunky" restyles the whole game, and Djinn refuses to ship a change that didn't visibly happen. Shipping is only half of it. Once the game is live, Djinn's LiveOps side keeps watching it. Every day it checks the game's players, economy and leaderboards and writes up what changed and what looks wrong. A studio can ask it anything in plain words ("why did yesterday's players spend less?"), and when something needs fixing, Djinn proposes the change, applies it after the studio says yes, and measures afterwards whether it worked. Every answer is checked by a second model before the studio sees it. Meta's models run through both halves. Muse Spark drives the builds end to end and does LiveOps' reading, answering and daily watch, and Muse Wonder independently checks every LiveOps answer.