AI MUD For Young Ladies; Or, How Do You Do Something With It?

For those of you who are just starting out, install Node.js and a coding agent.

If your machine is not powerful enough to run a local AI, ask your coding agent to help you connect to a small model through an API.

The first thing to do, to feel the power of a MUD, is ask your coding agent to create a minimalist one.

A few rooms, some doors, and objects you can examine and use.

Then create Ms. Thelma, an inhabitant who, when you first speak to her, asks you to connect an AI.

You configure the model’s address, and any required API key, in your local application.

Then Thelma gets to work.

She can add rooms, decorate them in the style of British royalty, and fill them with knickknacks and paintings.

Give her a little space to build in, and ask her to make something beautiful.

Next, connect an image-generation service, and ask your coding agent to give Thelma’s paintings an action.

When you commission a painting, it sends its description to the image generator and displays the resulting artwork in its frame.

In the lobby, the main room where visitors arrive, place a newspaper that draws from real news sources.

But make it a magical tomorrow’s newspaper.

It reads what is happening today, then imagines what tomorrow’s headlines might be.

You are creating objects in the MUD that represent something.

But they do not merely stand for things. They have state, and they have actions.

The painting holds its description and its artwork. The newspaper holds its sources and its latest edition.

You can examine them, use them, and ask an agent to work with them.

I am a huge fan of in-game pneumatic tube networks, which pass capsules from station to station.

Double-click a tube, and it shows you a flow graph: where the capsules go, and who is seated at each station.

To be seated means taking responsibility for the work that arrives there.

When a data capsule reaches the station, the seated agent receives an email.

These can be mailboxes inside the world. No outside email account is necessary.

I give my agents PDAs, so they have somewhere to read their mail.

But how do you use this?

Email the agent at the first station.

Speak in plain English. Explain what you would like made.

The agent tidies your request into a work order, places it in a capsule, and passes it down the line.

Each station contributes its part, and the capsule carries the work forward.

You can give objects actions, but the agent is your conversational API.

You do not have to memorize every command. You explain what you want, and the agent works with the objects to make it happen.

The world gives those requests somewhere to go, something to act upon, and somewhere to leave the result.

And when it is time to share, keep the working MUD on your own machine.

The agents work there. The API connections stay there. The credentials stay there, too.

Publish an explorable edition on GitHub Pages: the rooms, the descriptions, and all the art the world has made.

Present each room as a sideways-scrolling view of its walls, decorations, paintings, and doors.

Scroll far enough, and the walls wrap around to where you started.

Click a door to enter another room.

A little minimap can help visitors find their way around.

The published world does not need to keep generating. It lets people explore what has already been created.

They come to visit, not to configure an AI.

And to close this strange little tutorial with an interesting example:

Ask your pneumatic tube network to produce vivid, game-outro-style pixel art of breathtakingly beautiful animals that have never existed.

One agent imagines the creature. Another describes its impossible habitat. Another prepares the commission and sends it to the image generator.

The finished painting arrives in the gallery. ...