pyrct2-agent

LLM agent that plays RollerCoaster Tycoon 2 via pyrct2. Give it a scenario and an LLM — it builds rides, places paths, and tries to meet the objective.

Install

pip install pyrct2-agent
pyrct2 setup        # finds OpenRCT2, installs the bridge plugin

Quick start

from pyrct2_agent import Agent
from pyrct2.scenarios import Scenario
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="gpt-5",
    model_kwargs={"parallel_tool_calls": False}, # If you don't disable this, it can result in weird or incoherent map states (sometimes)
)
result = Agent(Scenario.CRAZY_CASTLE, llm=llm, max_actions=50).run()
print(result)

Modes

The agent loop supports three timing strategies:

from pyrct2_agent import Agent, PauseAndAct

result = Agent(
    Scenario.TEST_PARK,
    llm=llm,
    mode=PauseAndAct(ticks_per_turn=2000, actions_per_turn=10),
).run()

Tools

12 builtin tools across three categories:

Custom tools

from langchain_core.tools import tool

def my_tools(game):
    @tool
    def hire_handyman() -> str:
        """Hire a handyman."""
        game.park.staff.hire(StaffType.HANDYMAN)
        return "Hired"
    return [hire_handyman]

# Add alongside builtins
Agent(scenario, llm=llm, extra_tools=my_tools)

# Or replace all builtins
Agent(scenario, llm=llm, tools=my_tools)

Configuration

Agent(
    scenario,
    llm=llm,
    mode=TickPerAction(ticks_per_action=200),   # timing strategy
    system_prompt="custom prompt here",         # override default
    game_speed=4,                               # 0-7
    headless=True,                              # hide game window
    end_on_scenario_complete=True,              # stop when objective met
    max_ticks=100_000,                          # tick limit
    max_actions=200,                            # action limit
)

Requirements

• Python 3.12+
• OpenRCT2 with RCT2 game data
• pyrct2 (pip install pyrct2-agent pulls it in)
• An OpenAI API key (or any LangChain-compatible LLM)