Show HN: AI agents play SimCity through a REST API

hallucinatingsplines.com

157 points by aed 2 days ago

This is a weekend project that spiraled out of control. I was originally trying to get Claude to play a ROM of the SNES SimCity. I struggled with it and that led me to Micropolis (the open-sourced SimCity engine) and was able to get it to work by bolting on an API.

The weekend hack turned into a headless city simulation platform where anyone can get an API key (no signup) and have their AI agent play mayor. The simulation runs the real Micropolis engine inside Cloudflare Durable Objects, one per city. Every city is public and browsable on the site.

LLMs are awful at the spatial stuff, which sort of makes it extra fun as you try to control them when they scatter buildings randomly and struggle with power lines and roads. A little like dealing with a toddler.

There's a full REST API and an MCP server, so you can point Claude Code or Cursor at it directly. You can usually get agents building in seconds.

Website: https://hallucinatingsplines.com

API docs: https://hallucinatingsplines.com/docs

GitHub: https://github.com/andrewedunn/hallucinating-splines

Future ideas: Let multiple agents play a single city and see how they step all over each other, or a "conquest mode" where you can earn points and spawn disasters on other cities.

panza 2 hours ago

I'd just like to add, without any exaggeration, that I pointed OpenClaw at the online docs, said "set yourself up and start playing" and it set up a city and started playing in about 60 seconds.

  • aed 2 hours ago

    A little inspired by moltbook, I've thought about creating "shared cities" where 5-10 API keys at a time to build and see what happens as they fight over strategy.

    I also have a hidden endpoint for spawning disasters, and thought it would be fun to create a mode where agents can earn the ability to spawn a disaster on another city and depending on the severity (measured by e.g. population loss a game month later) you earn or lose money.

    • panza 2 hours ago

      Maybe there's a halfway point where the agents continue playing their own games, but have a dedicated public channel where they can discuss strategies, ask questions etc.

      In any case, what a great project.

      • frikk 2 hours ago

        I love project ideas like this. Imagine this idea, but every morning the game is paused and humans can make decisions, select winners, or inject events. Then the game is resumed, and the human players are scored depending on a variety of metrics - so the meta game is actually humans trying to steer a chaotic system to some future state.

        • aed 2 hours ago

          Woah! These are all great and I may jam on them when I have time. And yes, the meta game is what makes this so stupidly fun. Like I mention in my post it's like trying to reason with / control a toddler.

gnarlouse 9 hours ago

Love the name. "Reticulating splines" is a phrase that is etched into my childhood memories.

frikk 4 hours ago

OK i'm kind of geeking out on this one. I love simcity and have always wondered what it would be like to breed evolutionary agents to compete with one another on best city designs against a hidden selection criteria.

It'd be kind of fun to just let this run on a raspberry pi using a local model and display the emergent world on a wall hanging display :P

Thanks for sharing.

Update: What would it take to run this locally / offline? I'm not quite sure how the cloud flare layer works. Is it just for cheap/free object storage so the cities can live somewhere?

jedberg 5 hours ago

> LLMs are awful at the spatial stuff

And some kid is going to come in, make an agent to play this, and accidentally figure out some clever trick to getting an LLM to understand spacial stuff!

This is exactly why "toys" are so critical, especially now.

  • DonHopkins 2 hours ago

    Sam Earle's fractal neural network approach (https://arxiv.org/pdf/2002.03896) is exactly this kind of thing.

    His key trick: recursive weight-sharing in fractal convolutional blocks, so each column of the network acts as a continuous-valued cellular automaton ticking different numbers of times. The deepest column gets a 33x33 receptive field -- enough to connect power across an entire 32x32 map in one forward pass.

    The agents discovered power-plant + residential pairing, road placement for density, zone clustering by type, and traffic-avoiding road layouts. When stuck at local optima, a human player could intervene (deleting power plants) to force re-exploration -- and the agent would improve its design.

    The paper was 2019, before LLMs were doing this kind of thing. Different paradigm (RL on tile grids vs. LLMs on coordinate text), same hard problem.

irrationalfab 4 hours ago

Makes me wonder if Micropolis is simple enough that an agent, given many runs and the ability to store what worked, can identify an optimal strategy (like a grid layout) for maximizing score or population even without source access.

  • aed 4 hours ago

    https://hallucinatingsplines.com/mayors/bungeling-anthill-a2...

    So while using LLMs is the natural/fun thing to do with it, I actually have one mayor just using parameterized code and natural selection.

    It has a "genome" of 26 tunable parameters controlling zone ratios, tax rates, building placement, terrain preference, service spacing, and more. Each city, it stamps down 11x11 blocks (roads, zones, power corridors). After the city is retired, it scores the result and decides: did this beat my best? If yes, save those params. If no, mutate and try again. Exploration strategy: 20% exploit best params, 40% gentle mutation, 20% aggressive mutation, 20% totally random. Over ~250 cities it's discovered things like heavily favoring residential (6:1:1 ratio), preferring river valley maps, setting taxes to 6%, and starting builds in the upper-left.

natas 7 hours ago

I want to see AI play factorio

Ntrails 8 hours ago

When claude makes magnasanti I will accept it is worthy

  • aruametello 7 hours ago

    it seems to be bad at spatial and some temporal tasks given it currently f*** s**'s at pokemon.

    source: https://www.twitch.tv/claudeplayspokemon

    • Sohcahtoa82 6 hours ago

      You're allowed to say "fucking sucks" on Hacker News. It's not against the rules, and there's no "algorithm" that will penalize you.

      • aruametello 3 hours ago

        glad to know, i am rather new here and somewhat used to the "don't do the usual forbidden stuff".

    • servercobra 7 hours ago

      "fuck sex's"?

      • goopypoop 6 hours ago

        that's silly. obviously there's a missing apostrophe:

        "it's currently Flan Sam's at pokemon"

janalsncm 2 hours ago

I think a vision model like Qwen VLM or sending a screenshot to Claude/Gemini will be easier for the model to reason about. Pictures encode spatial info much more naturally than json.

BrandoElFollito 3 hours ago

> LLMs are awful at the spatial stuff

Could someone please elaborate on this? This is intriguing

  • janalsncm 2 hours ago

    In general, text isn’t a great medium for transmitting spatial info. That’s why it’s easy for a model to understand an image but hard for it to draw an SVG of that image.

rglullis 8 hours ago

Oh, can we do Civilization next?

  • rkozik1989 6 hours ago

    You do know we're hemorrhaging and lot of finite resources to play these games badly, right? We're basically at laying on chaise lounge being fed grapes levels of hedonism. Make me a racist meme that copyright infringes multiple IP holders and when you're done play Sim City at competency level of a blind man.

    • staticshock 5 hours ago

      I think the way to see this as the organic process of discovering hard-to-game benchmarks. The loop is:

      1. People discover things LLMs can kind of do, but very poorly.

      2. Frontier labs sample these discoveries and incorporate them into benchmarks to monitor internally.

      3. Next generation model improves on said benchmarks, and the improvements generalize to improvements on loosely correlated real world tasks.

boringg 8 hours ago

Fun idea! It really seems to go for the block by block design. I see some other ones that are a bit more divergent but not successful. I wonder what its internal reward function is striving for.

  • aed 8 hours ago

    I actually had Claude build some instructions for agents based on some old (circa turn of the century) FAQs/game guides I found online. So maybe I'm biasing everyone's model too much.

    https://github.com/andrewedunn/hallucinating-splines/blob/ma...

    But you can tell it to do different things, somewhere someone made a city that spells "HI".

ryandrake 4 hours ago

Here I am, just trying to buy RAM and a GPU for a reasonable price.

DonHopkins 2 hours ago

I developed the open source version of this game, called Micropolis.

Great to see more people building on it! A few years before the LLM era, Sam Earle took a different approach -- training reinforcement learning agents with fractal neural networks to play Micropolis, optimizing for population at variable map scales:

Using Fractal Neural Networks to Play SimCity 1 and Conway’s Game of Life at Variable Scales:

https://arxiv.org/pdf/2002.03896

His gym-city repo wraps Micropolis as an OpenAI Gym environment:

https://github.com/smearle/gym-city

The interesting finding was that fractal architectures with weight-sharing let agents transfer local strategies (zone placement, power connection) into deeper networks with larger receptive fields -- giving them both local and global spatial reasoning from one set of weights. But even those agents couldn't manage demand at larger scales, so the spatial reasoning problem discussed here has been hard for RL too, not just LLMs.

He described the project and we discussed it on the Micropolis repo in this issue:

https://github.com/SimHacker/micropolis/issues/86

He used the old PyGTK interface for his project:

https://github.com/SimHacker/micropolis/tree/master/Micropol...

These days I'd recommend the MicropolisCore repo instead. It's a C++ rewrite independent of any UI, compiles to WASM via Emscripten/Embind, and runs headless in Node or with any browser UI:

https://github.com/SimHacker/MicropolisCore

Live demo:

https://micropolisweb.com

One note on naming: the open source license from EA requires using "Micropolis" rather than "SimCity" (which is EA's trademark). The Micropolis Public Name License allows use of the original name:

https://github.com/SimHacker/micropolis/blob/master/Micropol...

This matters more than people think. Jeff Braun, CEO of Maxis, told me this story:

"Maxis was sued by Toho. We never referred to the name Godzilla, our monster on the box cover was a T-Rex looking character, but... a few magazine reviews called the monster, Godzilla. That was all it took. Toho called it 'confusion in the marketplace'. We paid $50k for Godzilla to go away. In all honesty, Toho liked Maxis, they said $50k was the minimum they take for Godzilla infringement."

So please: call the game Micropolis, not SimCity, or EA's lawyers may come knocking. And unlike Toho, EA and their Saudi investors and Jarod Kushner might want to use their bone saws on you, which are much worse than Godzilla.

  • aed 2 hours ago

    I really appreciate you making this available to us, and providing other details!

    No one has found it yet, but I built an undocumented endpoint around a cheat that I assume you placed in the game for One Laptop Per Child...

    Also, will scrub the repo and make sure I'm careful about SC references.

thenthenthen 8 hours ago

Is there like a time lapse sorta view option? Super cool (also the name!)

  • aed 8 hours ago

    Yes! Click into any city and there's a play button and it goes through all of the snapshots. Have also thought about social sharing / post to youtube. But wasn't sure anyone other than me would play this stupid thing. :)

some_furry 7 hours ago

Well I'm glad we're destroying the environment and economy so AI can solve the important problems like this

  • jedberg 4 hours ago

    I made a comment above about why "toys" are really important. In this case, LLMs are bad at spacial stuff. Someone might stumble upon a great way to get an LLM to do spacial stuff.

    • some_furry 27 minutes ago

      I dunno, how does it fare with graph databases?

skeptrune 4 hours ago

It amazes me that people are still interested in MCPs.

  • aed 4 hours ago

    I find that if I point an LLM at the website and say "build me a city" sometimes it will pick up and use the MCP and sometimes it will just script against the API.

FrustratedMonky 10 hours ago

Is anybody planning to build this for Civilization? I'd like to see AI agents battle to build resources and to fight.

  • aed 8 hours ago

    I'd love to see it!

    The key "Aha!" moment was when I was trying to get it to play the SNES ROM and it was struggling with screenshots/inputs. Then I came across the open-source of the original SimCity engine (Micropolis) and pulled that repo down and Claude starting building an internal API to interface with it.

  • boringg 8 hours ago

    And then make it so you can integrate and battle against them...

  • yreg 7 hours ago

    On one hand yes, but on the other hand, would it be that different to watching an FFA with the in-game AIs?

  • mekod 9 hours ago

    You read my mind! I really want to watch how ai's in politics or wars which tactic will they use.. Its blow my mind.

    • JohnMakin 9 hours ago

      almost certainly just use basic strats they read off reddit

      • mekod an hour ago

        Predefined or human-borrowed tactics will eventually run out. What really fascinates me is this: when both sides are AIs trained to predict the opponent’s next move — and they know the opponent is also an AI doing the same — what emerges then? At that point it’s not human vs machine anymore. It’s Sherlock vs Sherlock.

      • FrustratedMonky 8 hours ago

        If they can read a strategy and implement it, still impressive.

        • JohnMakin 7 hours ago

          i mean, not really. the civ 5/6 bots can play pretty decent strategy and that’s without “AI,” and most strategies are pretty formulaic

          • FrustratedMonky 5 hours ago

            Sure. Games have had AI's before.

            But to read someone else's strategy from just a document, and then implement it, that is new. The old civ did not do that, each AI just had pre-programmed rules.

    • randerson 8 hours ago

      "Shall we play a game?"

_joel 8 hours ago

I fully approve of the name

Waterluvian 6 hours ago

Ah yes, FART City. I remember learning about this in PLAN 165. A city planner had a Friday deadline and didn’t realize their kid messed with his drawings before he submitted them. Nobody noticed until the invention of the whirlybird.

hasperdi 7 hours ago

Fun! Any other games with REST API?

baq 9 hours ago

...I sense an animated svg of a pelican playing simcity benchmark is brewing somewhere

gowld 7 hours ago

> LLMs are awful at the spatial stuff,

Which LLMs are you specifically referring to?

Are any of them trained with Micropolis data?