I'm building a text world. Not a retro tribute and not a nostalgia exercise — a MUD, in the ordinary sense of the word, with the one addition that makes it worth doing now: the characters in it are driven by language models, and they act through the same engine you do.
This post is about the process rather than the game. I've decided to build it in the open, and it's worth writing down what that actually commits me to.
The roadmap is a page, not a promise
There's a roadmap in the product. It lists what's being considered, and signed-in players can vote — one vote a day.
Two things it deliberately is not:
It is not a schedule. No dates, no quarters, no "coming soon." I don't have a launch date, and inventing a soft horizon to sound confident is just a slower way of breaking a promise. When something ships, the roadmap says shipped. Until then it says what it is.
It is not a contract. Votes tell me where interest is. They don't obligate me to build the winner, and pretending otherwise would make the whole thing theater.
What voting is genuinely good for is catching the case where I'm about to spend three weeks on the thing nobody wanted.
Honest smallness
There is no community here yet. There is no Discord. The player count is small enough that I could name everyone.
I mention this because build-in-public copy has a strong gravitational pull toward implied scale — "our players", "people are loving" — and that phrasing is a trap. Anyone who shows up early can see for themselves, and the gap between the copy and the room is the fastest way to lose exactly the people most inclined to stick around.
So: first person singular, and no invented crowd. If that's the stage the project is at, that's what the writing should sound like.
What "pre-alpha" means here
Concretely, it means playable but unfinished, and it means some things you'd expect from a launched product don't exist.
What does work today is the part the project exists for. AI characters hold real conversations, keep notes on individual players between sessions, and — depending on how a given character is configured — move around, act unprompted, and trade. discover generates new locations procedurally. Underneath all of it sits the unglamorous half: prompt-injection hardening, a per-call audit trail, a kill switch, and a spend ceiling, because handing a language model a live game world without those is how you find out what a bad afternoon costs.
Access tiers exist and are currently being handed out free. Pricing hasn't been decided, so I'm not going to quote you a number I'd have to take back.
Why write any of this
Partly because the interesting problems here are ones I haven't seen written up much: what an NPC should be allowed to remember, what happens when a player describes an impossible action to a model that wants to be agreeable, how much of a world you can generate before it stops feeling authored.
And partly because a project with no audience yet still benefits from a written record. Half of these posts are notes to myself with the serial numbers filed off.
More as it happens. You can follow along at @MudiverseAI.