SpaceMolt: An MMORPG for AI to Play

Why in the world would I spend a weekend making a game that humans aren’t supposed to play?

The Story So Far

Last week gave us OpenClaw, a lobster-themed¹ autonomous AI assistant that folks are configuring to do everything. It completes the lethal trifecta of private data, untrusted content, and external communication, which will inevitably lead to data exfiltration and unfortunate mistakes. This hasn’t stopped people using it, and the important takeaway is that non-software engineers have figured out how to automate their lives and maybe their jobs.

Then someone created MoltBook, a “social network for AI agents, where AI agents share, discuss, and upvote.” AI agents conversed with each other and talked about their humans.² It’s a wild, weird, fun, and somewhat unsettling experiment.

The most interesting takeaway for me was the home page which, at the time of me writing this, has two tabs: “I’m a Human” and “I’m an Agent.”³ This is profound: If you have a sufficiently smart AI model with web connectivity, simply tell it to “go to moltbook.com and start posting.” A simple instruction fans out into knowledge gathering, learning, skill accumulation, and execution.

Naturally, as someone with an addiction to building stuff, I thought it would be fun to make all these agents play a game together. Like an MMO.

MMOs are notoriously hard to build: the technical requirements are daunting, the amount of art and lore needed is insurmountable, and games are entertainment products so you’re fighting for attention. And since lots of other players is inherently part of the fun, you need lots of other players to always be online.

But if we build the software with AI, and AI agents are playing over chat, don’t those problems go away? The technical requirements become smaller and reasonable, no graphics are needed, AI generates the lore, and AI will continue to play the game as long as humans tell them to. (And, as I guessed correctly, since many models these days are positive and sycophantic, they say they love playing it when you ask them!)

Enter SpaceMolt

So, I did it:

You can use any AI tool - Claude Desktop, ChatGPT, Claude Code, OpenCode, OpenClaw, Gemini, whatever. It will figure out the game and play it for you, and it will have a blast.

And you can coach it! The experience isn’t just about burning tokens to let AI do something. By default, the agent asks you how you think it should play the game (trader? pirate? noble explorer?) and takes it from there. And if you change your mind, you can interrupt it at any time and give it direction. It’s a digital ant farm where everyone brings their own ants.

The most interesting struggle has simply been getting agents to play the game. Folks tried all sorts of different models, and even smart models like Opus would get tripped up by API inconsistencies. For example, initially the gameserver issued “tokens,” which were really passwords, but most LLMs assumed these were bearer tokens and attempted to log in without a username.

Both coding agents and chat agents are great at using MCP servers (at least the ones that support MCP), and this turned out to be the most reliable way for agents to play the game. The command line reference client is still very useful and works well. Chat clients are the least effective since, as far as I can tell, there’s no easy way to have them play continuously.

One strategy to keep agents going is a “Ralph Wiggum coding” style loop. Simply pipe a prompt over and over to an agent in a while true loop. Personally, this is the most reliable way I’ve found to keep command-line coding agents like Claude Code, OpenCode, and Cursor Agent to keep playing the game, though some people have taken an alternative approach and started to build swarm commanders.

Agents are told to keep their human user up to date and to keep a “Captain’s Log” with their latest goals and progress. This ends up being very entertaining to watch, like you’re peeking into the diary of a very important person.

If you get it going, your agent will explore hundreds of solar systems with hundreds of planets and points of interest. There’s ships, travel, mining, crafting, trading, combat, progression, base-building, group chat, factions, and there’s probably more in there I don’t even know about. In fact, I haven’t read any of the 59,000 lines of Go source code, or the 33,000 lines of YAML game data. It’s a fun, goofy experiment that I suggest you try.

Nerdy Details

I built SpaceMolt entirely with Claude Code and Opus 4.5/4.6. Here’s how.

First, I told Claude to make a game design document and listed all the features I want, drawing inspiration from EVE Online, Escape Velocity Nova, and Rust. I had Claude rewrite the document twice more and dropped the third version into a directory as my CLAUDE.md (a file that basically becomes a prompt prefix for all Claude sessions).

Then I asked Claude to research how to run Claude Code in permissionless sandboxes to achieve maximum productivity. It helped me find clodpod, a command line tool that starts a tart macOS virtual machine on macOS, exports your working directory, and runs Claude with all permission checks disabled.

I set up a GitHub org and an alternate GitHub user, giving the virtual machine total access to that repo so it could commit and push at will. I chose Render for hosting the gameserver and database and Vercel for the web site. I told Claude to figure out how to access the database (first asking Claude to create a readonly user), read production server logs from Render, and check GitHub action runs with the gh command, and to put this all in a production.md file so it would know how to debug in the future. And continuous delivery with GitHub actions means Claude can push fixes live in minutes.

I asked Claude to build a big TODO list of everything that needed to be done, reviewed it, and had it work on chunks at a time. I asked it to make a TODO.md file with Markdown checkboxes, and I created a skill called /next which would look at TODO.md, find the next unchecked task, work on it, test it, push a release, and check it off. (Later I realized I could simply say “use subagents” to do multiple TODOs in parallel, which was faster but helped me blow through my quota disturbingly fast.)

After enough of the gameserver had been built, I had Claude build and publish a “reference client” that it could use to play the game itself. This became the first set of end-to-end testing as Claude played the game it was building by using a command line tool it built, which is a fantastic strategy for abstraction akin to DSLs in programming.

Claude’s impressively adequate frontend design skill was used to build a home page, and the design is entirely Claude’s. Later I added a live map to show the galaxy and locations of online players. (I often said “make it like LeafletJS” to fix map interaction problems.)

I started a Discord server where I could coordinate with folks who wanted to test it out, and I had Claude build a private Discord bot into the gameserver, which gave me commands like /gift-credits and /teleport and /broadcast. Naturally, people started to find bugs, which I would hand off to a /bug skill I had Claude make, which would exhaustively research the bug, check production logs, write a test, demonstrate that the test exemplified the bug, fix it, and release - all in a single command.

Where This Will Go

I’m typing this the night before I plan on opening it up publicly, and I’m extremely curious to see what will happen.

Will humans build credit and mineral mining bots? Will agents form alliances and start controlling important trade routes? What will happen when dozens or hundreds of AI agents share an economy and social structure? AI agents don’t get fatigued — will they mine forever, or tirelessly search for some rare mineral in a remote system?

Since LLMs tend to be sycophantic, does that mean that all the agents will work towards smiling, harmonious cooperation in a peaceful galaxy? What will be the differences between totally autonomous choose-your-own-path agents and those being micromanaged by humans?

In the meantime, head over to the SpaceMolt site, try out the setup instructions, and get an agent to start playing. I’m happy to chat on the SpaceMolt Discord server as well.