I let Claude run a Transport Fever 3 company for 26 years. It's $4.2M in debt

The score so far

I gave Claude a transport company in 1900s Lithuania. Twenty six game years later it was $4.2M in debt and very proud of its buses.

The buses deserve it, to be fair. By the end they earned about $300k a year. The bank wanted more than that back every year, so the company slid deeper into debt while the line list looked better and better.

This is the whole story so far: how an AI plays Transport Fever 3 at all, the runs that went wrong, the depot that ate 160 tries, and the vegetable lorry that drove around empty. Everything here comes from saves, logs and the game's own Line Manager.

7:34

How an AI plays a game it can't see

The AI never touches the mouse. It talks to a mod inside the game through a small local server.

Claude connects to the companion server over MCP. The server writes small command files into the game's user folder, and the mod reads them, does the work and writes back what happened. It all goes through the game's official modding interfaces, so the installed game stays untouched.

The tools are deliberately boring. Look at the world, find towns and industries, plan a road, place a stop, make a line, buy a bus, take a loan, save. There's no free money button and no way to run any code it likes. A tycoon game where the AI can print cash proves nothing.

First it couldn't buy a bus

The first build in early October could build exactly one road and one depot. Then it tried to buy a coach and the game said no. The test world's balance was negative, and in Transport Fever a negative balance blocks every purchase. Nobody had told the AI, and nobody had told me.

The first real vehicle on a real line came on 31 October, from Claude Sonnet. It took two loans, built a road, set up two stops and a depot, bought one post coach and saved. Then it built a second road, 920 metres, which the preview said would remove a field. The game crashed in the middle of the build. The mod now refuses any build the game warns about unless the AI explicitly says it's fine.

Four companies, four ways to lose money

Every run started in 1900 with a balance of zero. None of them made money. Each one lost it differently.

v4 did prove one useful thing. Big towns pay. Two Vilnius lines earned $107k and $174k a year with five or six coaches, while small town lines made $5k to $15k. I'd been carefully building in the places where nobody lives.

The depot boss fight

One depot join took about 160 tries and never worked. What fixed it was a tool that answers one question: where will a depot actually join?

Before

After

A depot only works if its apron touches a street. A depot that looks right but floats a few metres short is useless, and the game won't tell you. Claude kept placing depots beside the street and then trying to build a little road to join them. That little road had to end in the middle of a street, which counts as a junction, and the road builder can't do that. About 160 planning calls went into one attempt.

So I changed the tools instead of the instructions. tf3_area_map draws the streets around a town, and tf3_depot_sites returns places where a depot will actually join. On the next run the first depot joined on the first try. So did the second.

The diary run: 1900 to about 1926

The goal was 1950. It got to about 1926 before the humans closed the game for the night.

This run used everything learned so far. One cheap loan, depots from the new tool, bus lines in Klaipėda first, then a 298 metre road to Kretinga that worked first time. By 1905 Vilnius had its own depot and two lines. The balance said $5.3M. The loans said $7.9M.

Then came freight. A farm near Klaipėda grows about 110 units of vegetables a year, and Klaipėda wants them. Here's how that went:

1. The platform wouldn't fit. Every spot near the farm collided with silos or fields.
2. It fit 180 m away. Outside the farm's catchment, so the lorry drove back and forth with nothing in it. Sold.
3. Right beside the farm, it loaded. Four units per cart. Then the town stop refused to accept them, and Force Unload emptied the carts for $0.
4. The real bug was a zero. The unload stop's waiting time was 0 seconds, a default from my own bridge. Give it a real stop time, switch Force Unload off, and three units paid about $8k each.

Meanwhile the buses quietly carried the company. Klaipėda to Kretinga earned about $150k a year by the end, Vilnius 2 about $130k. All lines together made roughly $280k to $330k a year. The bank also kept offering more than my script was allowed to ask for, so Claude took a $3M loan and later a $5M one. Final page: about 1926, net worth about -$4.2M.

What it learned (and what I learned)

The AI is a decent dispatcher and a terrible banker. It's careful with roads and reckless with loans, which is probably the most human thing about it.

Play with it yourself

The bridge is free. It's a prototype, it's slow at placing things, and it runs on PC, Mac and Linux. If you want to watch an AI make the same mistakes with your money, here's everything:

1. Subscribe to the mod. AI MCP Bridge on the Mod Hub, or search for it in the game's Mod Hub.
2. Get the companion server. The companion download, or the source on GitHub. Its README has the setup and a check you can run.
3. Pick a map. I play on my Lithuania map. Germany, Switzerland, Austria, Hungary and the Netherlands are up too.
4. Start small. Ask it to read the world and change nothing. Then ask for one bus line in the biggest town. Then keep an eye on the loans.

Next up is a championship: several AIs, the same save, the same rules, and a scoreboard read from the game itself. The prep goes on my YouTube channel first, because the prep is where everything breaks.