Agents In Use
What’s new todayLast update
— use cases added today

Claude builds a Micropolis city through an API

Andrew Dunn reports getting Claude to build cities in Micropolis after a failed attempt with an SNES SimCity ROM, with persistent public city histories.

OUR ADAPTATION

Suggested prompt

Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.

Use the Micropolis API to build a small city with power, roads and room to expand. Inspect the simulation after each development phase, explain disconnected areas, and return its public history link for review.

Evidence at a glance

Sourced account
Source author
aed / Andrew Dunn
Original account
Read the source ↗
Published
9 Feb 2026
Source link checked
15 Oct 2026
Reported result
Dunn reports a working Micropolis integration after the failed ROM approach, while noting poor spatial decisions involving roads, buildings and power.
Human involvement
The builder changes the simulator interface and observes layout weaknesses; the account describes human attempts to control scattered buildings and poor connections.

A linked account supports attribution. It does not mean we independently reproduced its result.

Report an issue with this caseOpen correction email

Opens your email app with the case link included. Nothing is sent automatically.

How the agent did it

  1. Attempt the city task with an SNES ROM and encounter difficulty.
  2. Switch to Micropolis and expose city simulation operations through an API.
  3. Let Claude operate the city through tools and inspect its public history and layout mistakes.

Andrew Dunn first tried to have Claude play an SNES SimCity ROM. That approach proved difficult, so he switched to Micropolis, the open source city simulation engine, and added an API. His firsthand account says this change got the agent working and grew into Hallucinating Splines, a platform with a separate persistent simulation for each city.

The task is to give an agent a city building goal and let it operate the simulator through tools. The platform offers both REST endpoints and an MCP server for clients such as Claude Code. Cities stay public and can be inspected with their recorded histories, so the output is an evolving simulation rather than a textual description of a proposed city.

Dunn also reports a concrete weakness. The models scatter buildings and struggle to connect roads and power lines. A working API therefore did not solve spatial planning. The present site displays city populations, scores and year-by-year views, but those changing totals are not attributed to the original Claude run. The case supports an inspectable city building experiment, with human steering around layout mistakes, rather than a claim that the agent built an efficient or successful municipality.

Prompt
Suggested prompt (written by Agents In Use, not the author's)

Source: Hacker News · aed / Andrew Dunn

Frequently asked questions

Why switch away from the SNES game?

Dunn says he struggled with the ROM approach and got the workflow working after adding an API to Micropolis.

Can people inspect the resulting cities?

The platform makes cities public and records their histories.

What did the agents struggle with?

The author reports scattered buildings and difficulties with spatial layout, roads and power lines.

More details

About this account

Claude operates a headless city simulation through API tools; persistent city state and public histories expose the consequences of its building choices.

Firsthand account opened and read. Firsthand HN story API timestamp 1770652473 gives 2026-02-09. Outcomes are author-reported and were not independently reproduced. Suggested prompt written by Agents In Use. Live site metrics are not used as the outcome of the original Claude run.

Credit: aed / Andrew Dunn

Source published 9 Feb 2026

Source checked 15 Oct 2026

Tools and topics

Micropolis · REST API · MCP server · Cloudflare Durable Objects

Games · Claude Code

Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.