rct2-agent
Enables an agent to sidecar-manage an OpenRCT2 park while you play: reading park state, adjusting prices, staff, marketing, loans, taking screenshots, and controlling game time, with management-only actions that never build.
README
rct2-agent
An MCP server + OpenRCT2 plugin that lets an agent sidecar-manage your park while you play: read the state, act on the business (prices, staff, marketing, loans), see the park (screenshots), and control time. Management only — the agent tunes the business and can recommend building, but never builds.
How it fits together
Claude Code (WSL)
│ stdio (MCP)
▼
MCP server ── dist/server.mjs, run by Windows node.exe
│ TCP 127.0.0.1:7860 (newline-delimited JSON)
▼
Plugin (intransient) ── loaded inside OpenRCT2, LISTENS on the port
│
▼
OpenRCT2 game
The server runs under Windows node.exe (there is no node in WSL), so both
the server and the plugin sit on the same Windows localhost — no WSL↔Windows
networking involved. The plugin listens (it's intransient, so it stays alive
across park loads); the server dials in and reconnects as needed.
Requirements
- OpenRCT2 0.5.4+ (QuickJS engine — needed for Promises/ES2023).
- Windows Node.js (found here at
C:\Program Files\nodejs\node.exe).
Build & install
From WSL, using the Windows toolchain:
# helper wrappers (or just call node.exe / npm-cli.js directly)
NODE="/mnt/c/Program Files/nodejs/node.exe"
NPM=("$NODE" "C:\\Program Files\\nodejs\\node_modules\\npm\\bin\\npm-cli.js")
"${NPM[@]}" install
"${NPM[@]}" run build # builds dist/rct2-agent.plugin.js + dist/server.mjs
"${NPM[@]}" run install:plugin # copies the plugin into your OpenRCT2 plugin folder
install:plugin copies to
C:\Users\casey\OneDrive\Documents\OpenRCT2\plugin\rct2-agent.plugin.js
(override with RCT2_PLUGIN_DIR).
Wire up the MCP server
A ready-to-use project config is in .mcp.json. In Claude Code, from the project
directory, it will be picked up automatically (approve it when prompted), or add
it explicitly:
claude mcp add rct2-agent --scope project \
-- "/mnt/c/Program Files/nodejs/node.exe" \
"C:\\Users\\casey\\OneDrive\\Desktop\\dead code projects\\rct2-agent\\dist\\server.mjs"
Run
- Launch OpenRCT2 and load a scenario. Confirm the plugin loaded — the in-game
console shows
[rct2-agent] listening on 127.0.0.1:7860. - Start Claude Code in this folder. The MCP server connects on first tool call.
- Ask the agent to manage the park. The core loop is measure → act → wait →
measure: read baselines, make changes,
advance_days(7), re-measure.
Tools
Read (eyes): get_park_summary, get_finance_report, list_rides,
get_ride, list_shops, get_guest_overview, sample_guest_thoughts,
list_staff, get_scenario
Act (hands): set_ride_price, set_shop_price, set_park_entry_fee,
open_ride, close_ride, set_inspection_interval,
start_marketing_campaign, set_research_funding, hire_staff, fire_staff,
set_staff_patrol, set_loan
See (vision): capture_view, capture_ride, find_location
Time: get_clock, set_game_speed, pause, resume, advance_days
Save: snapshot, list_snapshots
Notes on behavior
- Money is in plain dollars at the tool boundary; internally OpenRCT2 stores
tenths, converted in
src/shared/protocol.ts(MONEY_FACTOR). If a value looks off by 10×, that's the one knob to check. - Coordinates in
capture_view/set_staff_patrol/find_locationare in tiles (converted to map units internally). advance_daysunpauses, runs at max speed, and auto-pauses when the target date is reached — then returns the new clock. It blocks until done.- Snapshots save to
save/agent__<play>__<label>.park, flat in the save root so they appear in the in-game Load Game menu —context.saveGamesilently refuses any filename with a path separator, so subfolders are not possible. Restoring is a manual Load in-game — the plugin API can save but not load a park. - Writes go through game actions, so the game's own limits apply (e.g. the ride price cap). A rejected action comes back as a tool error.
Dev
npm run typecheck— tsc, no emit.node scripts/smoke.mjs— loads the built plugin into a stubbed game and exercises the handlers over a real socket.node scripts/mcp-smoke.mjs— spawns the built MCP server and drives it with real MCP JSON-RPC against a fake plugin.
Layout
src/shared/protocol.ts wire protocol + money conversion + method names
src/plugin/main.ts the intransient in-game plugin (TCP listener + handlers)
src/server/rct-client.ts reconnecting TCP client to the plugin
src/server/index.ts MCP server: tool definitions -> plugin methods
esbuild.mjs builds both bundles
scripts/ install + smoke tests
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。