mythrun — to be continued?
Mythrun was my attempt to build an AI Dungeon Master for D&D 5e that I'd actually want to play with. I've been a DM forever, and the existing AI offerings weren't getting some pretty basic things right. The lack of a traditional session structure made play sessions feel rambling and structurally incoherent. The models were too willing to let players do things their characters weren't capable of. And an AI chat interface is a bad fit for D&D combat, which should play almost like a traditional video game.
My idea was to have the AI do what a good DM does: prep sessions ahead of time, keep track of the world, and hold firm to the rules. The roleplaying happened through chat, but when combat started, the game switched to a tactical UI with a deterministic rules engine and AI-generated voxel battlemaps. I'm particularly proud of the battlemaps; voxels are great for D&D combat, and I'm surprised that there aren't more VTTs trying this.
Underneath all of this was a fairly involved orchestration setup powered by Claude. Different models handled storytelling, planning, and bookkeeping, with explicit limits on what each could see and do. Dice rolls and rules were handled by deterministic code.
One of the more interesting architectural moves that I came up with was something I called “enforced ignorance.” The storyteller didn't have access to the planned plot. If an AI knows that an NPC has a plot hook to share, it tends to barrel straight towards it, making the conversation feel stilted and scripted. Instead, a higher-level overseer watched the conversation and waited until enough rapport had been established before giving the storyteller what I called “stage directions”: something like “NPC X knows more about Y, lead the conversation in that direction.” The idea was for an NPC to feel like a person who happens to bring up something interesting for you to do.
Unfortunately, keeping the narrative quality where I wanted it meant that a two-hour session cost almost $20 in API calls, even after extensive cost optimization. I couldn't make that work as a product, so I mothballed Mythrun in August 2026. I plan to open source some of the more interesting pieces of the architecture for anyone in the D&D community who could get some value out of them.