Jev Plays Pokémon Red: an AI model plays the game live, every decision on screen

Christian Mathiesen, a developer at Frigade, has published a live demo in which Jev, TypeSafe AI's decision model, plays through Pokémon Red. The page streams the game from YouTube, and a panel on the right shows every decision the model makes together with the odds it gave each option.

How the harness works

The core is a Game Boy emulator written in Node. It reads the game's RAM and turns it into a state description: the current map, the party, battle data and the on-screen text. From that state the harness builds a list of legal options, adds facts such as type matchups and damage estimates to each one, and asks Jev to pick. The choice is then turned into button presses, with A* pathfinding for movement and separate routines for dialogs and menus.

Jev answers every menu question, from names and the starter to shops and moves, decides what to focus on next, which Pokémon to catch or swap at the PC, where to go and who to talk to, and every battle action: move, switch, item, ball or run.

What the harness does not do

The harness only reads memory and presses buttons; it never writes to game memory, so no scripts or cheats are involved. Game knowledge is limited to story milestones, meaning the next goal and where it happens, and progress is checked against the game's real event flags. Hidden items are not shown, since a human player would not know where they are either.

House rules and loop protection

Text speed is set to FAST and battle animations are off, once at boot in the Options menu. The player is named JEV and the rival BLUE, and every caught Pokémon gets a made-up nickname that Jev spells one letter at a time. Options already tried without anything changing get tagged, the harness samples an alternative if the same failing choice keeps returning, and as a last resort it reloads the latest milestone checkpoint.

Code, cost and legal notes

The source is on GitHub under GPL-2.0 and requires the user's own legally obtained copy of Pokémon Red; no ROM is included or distributed. Every call is logged with its state, options, probabilities and latency. The README puts Jev at $0.042 per million input tokens with free output, or roughly $1 to $1.70 per day at the observed 800 to 1,300 calls an hour. The project is a fan experiment, not affiliated with Nintendo, Game Freak, The Pokémon Company or TypeSafe AI.

The demo also doubles as a pitch for Frigade, described on the page as an AI assistant that learns a product on its own and shows each user the next step inside the app.