ChainCraft Is Building a Multiplayer Game Platform That Uses AI Only When It Has To

Ryan Beltran went into "Superhero Showdown" with a strategy. He had noticed through repeated play that ChainCraft's AI narrator showed reluctance to eliminate cute characters. So he submitted his fighter as "a super cute puppy with big eyes and an even bigger heart." His co-founder Eric Wood, not knowing what Ryan had chosen, entered as Schrodinger's cat: neither alive nor dead, and therefore undefeatable. The AI announcer, configured to sound like a wrestling commentator, ruled that the puppy's love attacks "pass harmlessly through Schrodinger's cat." Eric won. "This is the first time that I've lost with this strategy," Ryan said afterward.

That live demo, run during a DevNTell appearance in early 2026, illustrates what ChainCraft has been building since the two co-founders first appeared on the show: a multiplayer game platform that uses AI for specific, bounded tasks like narrating battle outcomes, while keeping everything else deterministic.

From Discord to the Questbook Grant

When Ryan and Eric last appeared on DevNTell, ChainCraft was running inside Discord and had just received a Questbook grant from the Arbitrum ecosystem. By the time of this episode, that grant period had ended. "It's kind of a bookend to where we're at," Ryan said.

The company has a third co-founder, Matt Daly, and a front-end and smart contract contributor Ryan refers to as Mark. Over the intervening period, the platform moved from its Discord roots to a mobile-first, browser-based web app. Each game a user creates generates a cartridge graphic and mints as an NFT, establishing the creator as the owner of that game specification. Profiles are handled through Thirdweb. Games are multiplayer by default. The platform also added token duels, where players can stake a wager on a two-player match, and item creation, where completing a game can trigger the minting of a character token that persists across sessions.

At the NYC Open House, ChainCraft added the ability to pull live data from blockchain sources or REST APIs. One demo, which Ryan nicknamed "Tokimon," tied a character's strength to the current price of ETH or Bitcoin. These features, Eric noted, slotted into the existing architecture without major rewrites, which he described as one of the more satisfying signs of how the system has matured.

As Little with AI as Possible

The most significant architectural change since the last episode is not a new feature but a deliberate narrowing of where AI appears in the stack. Eric summarized the guiding principle plainly: "our strategy is to do as little with AI as possible."

In earlier versions of ChainCraft, a single large language model was responsible for tracking and updating the entire game state. The problems that came with that approach were predictable: the model would hallucinate, forget details, or make decisions inconsistent with prior game events. To address this, Eric introduced a concept he calls "artifacts," which separate the game schema, state transitions, and instructions into distinct components. Rather than writing the state directly, the AI now produces what Eric calls "state delta ops," a set of proposed changes that the system then applies through deterministic code. The AI tells the engine what should change; the engine decides how.

A second optimization, narrative compression, addresses the cost problem. Text-based game specs tend to accumulate lengthy narrative descriptions of characters and rules. Those descriptions are expensive to include in every prompt. ChainCraft now compresses those narratives to markers at build time, expanding them only when a human is reading or when the engine needs to generate output for a player.

When artifact generation fails validation, the system now attempts to repair itself. A second model receives the failed artifact, the validation error, and instructions to make targeted corrections. "The game engine's much more reliable," Eric said, describing the outcome of these combined changes. "We see no variability between game sessions."

The most recent evolution is experimenting with fully deterministic mechanics for rule resolution. Something like rock-paper-scissors, which previously required an LLM to interpret the outcome, can now be handled by a generated function that runs the logic without calling a model at all. The goal is a game engine that reaches out to an LLM only when it genuinely needs one, for narrative generation or adjudication of ambiguous situations.

An Underserved Niche

ChainCraft was not the founders' first concept. Eric recalled that early conversations included D&D-style games, but he felt the narrative generation requirements were too demanding. "I said, oh, that requires a lot of narrative work, and narrative generation is just not something that AI is great at right now," he explained. The company settled on text-based games, card games, and trading card game mechanics, a category Eric describes as "an underserved niche in game content creation."

That focus has shaped the product roadmap. The current interface is text-based, which places cognitive demands on players to track game state in their heads. Eric said the team is working toward visual interfaces, with plans to generate UI descriptions alongside game specifications so that different rendering targets, including Phaser, Babylon, or eventually Unity or Unreal, could interpret the same underlying game without a separate build process.

The community layer is also part of the design. ChainCraft's spectator mode allows anyone to watch a live game session. Users can remix another creator's game or invite someone into the builder to give feedback in real time. Ryan compared the long-term direction to itch.io: easy sharing, in-app play, and a growing set of tools that compound on each other over time.

Creativity as the Actual Product

Ryan has a background in photography, including darkroom work, and watched digital photography arrive during a period when practitioners in both camps were openly skeptical of each other. He drew a direct line between that moment and the current debate about AI in games and art, describing AI as another tool that artists can choose to pick up or not.

Eric framed the question in terms of reach. He described thinking about people who have genuinely good creative ideas but lack the time, resources, or technical background to turn them into a playable experience.

"I look at what we're doing as expanding the scope and democratizing that creativity and giving people an opportunity who have great ideas but don't have the resources. We all get to experience their great creative ideas where we couldn't do that if we didn't have tools like this."

For aspiring founders watching the episode, Ryan pointed to non-dilutive funding, specifically Arbitrum's grant programs through Questbook, as a starting point worth pursuing carefully. He recommended putting real time into writing a grant proposal and noted that Arbitrum will provide feedback even on rejected applications. Eric's advice focused on sustainability: "Don't skimp on your go-to-market," he said, adding that clarity about audience, monetization, and what personal sacrifices are actually sustainable matters before the first line of code gets written.

ChainCraft is currently accepting play testers through its Discord, with a broader public launch planned once additional features are in place. As for the Schrodinger's cat ruling, Ryan seemed to accept it: a character that cannot be definitively observed, it turns out, also cannot be definitively defeated by a puppy.