Codex Mortis 1.0: the AI-made game that found its players
The roguelite its studio bills as the first fully AI-developed game has left Steam Early Access, adding four-player online co-op, ten schools of magic and 200 spells.
Codex Mortis left Steam Early Access today, taking with it a year-old argument about whether a game built entirely by artificial intelligence can survive contact with players. Version 1.0 shipped on 24 September for Windows, Linux and Steam Deck at $3.99, currently 30% off for launch, and it arrives with four-player online co-op, ten schools of dark magic, 200 spells and 31 heroes. CRUNCHFEST, the studio publishing it, bills it as the first fully AI-developed game to reach release.
Strip away that angle and it is a shape you already know: an isometric horde-survivor bullet hell whose roguelite loop runs on a spell deckbuilder. You play a necromancer, build a deck from the spells you own before each run, then carry it into a battlefield of the undead. Points earned mid-run buy booster packs, every pack widens the pool for your next deck, and each school of magic mutates the others, so a Harvest build plays nothing like one leaning on Blood Magic. The game has been in Early Access since 26 March.
What 1.0 actually adds
The full release is a substantial one. Five schools join the original five — Harvest, Death Knight, Spiritism, Desolation and Apothecary — taking the spell pool to 200 and the roster to 31 playable necromancers. Four new Apocrypha levels open up, ending in a three-phase boss with 600,000 hit points, and all of them run in Survivor mode as well. The online co-op promised on the first day of Early Access has finally landed: up to four players can host or join over Steam, browse friends' sessions, lock a lobby with a password or share its code, and rejoin a run in progress after a crash. Local co-op stays. The game is now fully isometric, supports ultrawide screens and controllers, is Steam Deck Verified and ships in 28 languages.
One developer, one model, 350,000 lines of TypeScript
The 15-minute documentary released alongside 1.0 is the more interesting artefact. GROLAF's first attempt on an off-the-shelf engine stalled, so he asked the model which programming language it knew best, took the answer, and built a custom engine the AI could read and run entirely through text. The project runs to roughly 3,450 commits and more than 350,000 lines of code since July 2025. He also wrote an instruction file that opened by telling the AI its user was already irritated after repeated failures: stop explaining, stop apologising, get on with the work. His verdict in the film is that it helped.
GROLAF's own framing is a lot less romantic than the coverage. “I still use my experience as a lead developer every day,” he says in the release. “Define the task, delegate it, review the result. I'm still leading the work. The contractor has changed.” On the slop charge his answer is blunter still. “We've had asset flips, Unity slop, Unreal slop. To me, the problem is lazy execution. The tool alone doesn't decide the quality. Like a typewriter or a car, AI is a tool. What we make is up to us.”
From “slop” to a Steam listing that sells
The free demo arrived on 9 December 2025 and was met with a headline — “a milestone for slop everywhere” — that has followed the game ever since. It also found an audience: 10,500 demo players averaging 71-minute sessions, a run to 322 concurrent players on SteamDB after the streamer Asmongold played it on camera, and more than 20 community patches through Early Access. Steam now lists Codex Mortis as Mostly Positive across 78 user reviews.
Our opinion
Codex Mortis is a useful test of the AI question because nothing about it is spectacular. A $3.99 survivors-like with an isometric camera and a deck-building screen sits in the least prestigious corner of Steam, which is exactly where the argument gets settled: not in talks about what models will do next, but in whether 78 people who owed the game nothing paid up and said something kind. The claim worth taking seriously is narrower than the headlines suggested. It is not that a model wrote a game; it is that one person kept the job of lead developer and handed the subcontracting out. “Slop” was a prediction about volume, and that prediction was correct — cheap tools do fill the store. What it missed is that volume was never the hard part. Anyone who has shipped anything knows the hard part is deciding what not to build, and no model has sat through that meeting yet.
The other thing the pile-on got wrong is the scoreboard. The mechanic of the outrage cycle is that the first trailer gets a million impressions and the 1.0 patch notes get none, so the loudest opinion anyone holds about this game was formed from a 90-second video rather than a run through the Apocrypha. That is not a defence of the game's quality — 78 reviews is too small a sample to settle it either way — but it does mean the verdict on AI-made games is currently being written by people who have not played one.