openrct2-mcp
Bridges a running OpenRCT2 game to AI agents, enabling them to read park state, optimize parks, and build roller coasters.
README
OpenRCT2 MCP AI Park Control
Bridge a running OpenRCT2 game to Cursor (or any MCP client) so AI agents can read park state, optimize parks, and build roller coasters.
Quick links: License · Issues · Third-party notices
Prerequisites
- OpenRCT2 0.5.x with RCT2 game files installed
- Python 3.11+
- Node.js 18+ (plugin packaging only)
This repo was tested against OpenRCT2 v0.5.1 on macOS.
Quick setup
# 1. Install Python deps + openrct2-bridge plugin
chmod +x scripts/*.sh
./scripts/install-bridge.sh
# 2. Install ride-builder coaster plugin
./scripts/build-plugins.sh
# 3. Launch OpenRCT2, load a park save
# Check the in-game console for:
# - Bridge port (starts at 20020)
# - Ride-builder port (starts at 20021)
# 4. Verify connections (with OpenRCT2 running and a park loaded)
./scripts/check-connection.sh
# Optional: automated headless smoke test
.venv/bin/python scripts/verify-e2e.py
Cursor MCP
Project MCP config lives at .cursor/mcp.json (portable template: .cursor/mcp.json.example). Paths use ${workspaceFolder} and set PYTHONPATH to mcp-server/ so python -m openrct2_mcp resolves without a global install. Copy or adapt the example if your editor does not expand ${workspaceFolder}.
Reload MCP servers in Cursor after OpenRCT2 is running with a loaded park.
Note: End-to-end tools (live park mutations, vision capture, coaster placement) require a running OpenRCT2 instance with plugins loaded. CI runs offline unit tests only.
Example prompts:
- "Use openrct2_status, then get_park_overview and suggest three improvements."
- "List rides sorted by excitement and close the worst performer."
- "Build a small wooden coaster: coaster_create, then loop coaster_get_valid_pieces / coaster_place_piece until the circuit completes."
Architecture
Cursor AI --MCP--> Python MCP server --TCP--> openrct2-bridge (park ops)
`--TCP--> ride-builder (coasters)
| Component | Role |
|---|---|
| openrct2-bridge | 82 game actions, 77 state queries via pyrct2 |
ride-builder.js |
Coaster-specific TCP API (valid pieces, place track, test ratings) |
mcp-server/openrct2_mcp |
Curated MCP tools wrapping both plugins |
MCP tools
Park optimization
| Tool | Description |
|---|---|
openrct2_status |
Connection health for bridge + ride-builder |
get_park_overview |
Rating, guests, cash, scenario objective |
list_rides |
All rides with ratings and profit (fast per-ride queries) |
get_ride |
One ride by id |
get_park_messages |
Complaints and awards |
get_guest |
One guest by entity id |
list_guests |
Guest count by default; full_scan=true for samples (slow) |
place_flat_ride |
Place flat rides (e.g. gentle.MERRY_GO_ROUND) |
place_stall |
Place stalls (e.g. stall.DRINKS_STALL) |
manage_paths |
Paths, queues, benches, bins |
manage_staff |
Hire staff, list, set patrol/orders (by staff_id) |
list_staff_tool / set_staff_patrol_tool |
Per-staff patrol management |
assign_staff_to_path_corridor_tool |
Patrol box covering path tiles |
optimize_staff_coverage_tool |
Hire + assign staff along main paths |
park_health_report_tool |
Complaints, low satisfaction, recommendations |
get_map_bounds_tool / get_map_region_tool |
Park-scale spatial grids |
get_path_graph_tool / find_buildable_loop_tool |
Path graph and build surveys |
apply_theme_preset_tool |
Theme along paths (see themes/cute.json) |
place_small_scenery_tool / paint_terrain_tool |
Decoration and terrain |
set_ride_price / open_ride / close_ride |
Ride operations |
refurbish_ride_tool |
Renew a ride (resets age/reliability; closes and waits for guests by default) |
set_park_settings |
Entrance fee, open/close park |
advance_time |
Step simulation ticks |
get_research |
Research state |
Coaster building
| Tool | Description |
|---|---|
coaster_create |
Start a new coaster shell |
coaster_get_valid_pieces |
Legal next track pieces |
coaster_place_piece |
Place validated track (z_is_train_entry for endpoint Z) |
coaster_place_next_piece |
Place next piece from valid endpoint (Z-safe) |
coaster_build_perimeter |
Build flexible perimeter loop along guide ring (detours around obstacles) |
coaster_obstacle_map_tool |
Perimeter obstacle heatmap (paths, track, scenery, slopes, unowned ring stats) |
coaster_plan_route_tool |
Dry-run simulate route (probe+undo, no permanent build) |
coaster_plan_ahead_tool |
Beam search 2–4 pieces ahead |
coaster_prepare_corridor_tool |
Buy/clear/flatten along guide corridor (dry_run default) |
coaster_track_context_tool / coaster_full_track_context_tool |
Endpoint context with height + collisions |
get_tile_height_tool / coaster_site_survey_tool |
Ground tile_z, slope, flat build sites near x,y |
coaster_height_context_tool / coaster_probe_piece_tool |
Z vs ground; per-piece collision probe |
coaster_start_track_tool |
Create coaster + BeginStation at surveyed tile |
coaster_rank_next_pieces_tool |
Rank pieces by clearance, height fit, and collisions |
coaster_undo |
Remove last piece |
coaster_finish_station |
Place entrance/exit |
coaster_test |
Test ride and return ratings |
coaster_stats |
Current excitement/intensity/nausea |
coaster_delete |
Demolish coaster |
list_loaded_ride_objects |
Valid ride object indices |
Coaster workflow
list_loaded_ride_objects— pick a validride_objectindexcoaster_site_survey_tool(x, y)— find flattile_zand nearby obstaclescoaster_start_track_tool(x, y)orcoaster_create+ manual station placement- Loop:
coaster_rank_next_pieces_toolorcoaster_probe_piece_tool→coaster_place_next_piece - When circuit completes:
coaster_finish_station→coaster_test
Use get_tile_height_tool for a single tile; coaster_height_context_tool for endpoint Z vs ground.
Perimeter coaster workflow
coaster_obstacle_map_tool(inset=18)— survey ring clearance and unowned tiles on the guide ringcoaster_plan_route_tool(inset=18)— simulate trace without building- Optional:
coaster_prepare_corridor_tool(dry_run=true)then apply withconfirm_destructive=true coaster_build_perimeter(dry_run=true)→coaster_build_perimeter— execute build
Terminal debug (no MCP): .venv/bin/python scripts/plan-coaster-route.py --simulate --inset 18
Vision (screenshots + spatial context)
The MCP server can see the park window while also returning structured map data.
| Tool | What it does |
|---|---|
capture_game_view |
Screenshot of the OpenRCT2 window (returned as an MCP image) |
inspect_area_at_tile |
ASCII path/queue grid + nearby rides + optional screenshot |
Recommended build loop:
inspect_area_at_tile(39, 44)— Triple Corkscrew entrance area (example)- Plan paths / queue / ride placement from ASCII map + screenshot
- Apply changes via
manage_paths,place_stall,coaster_*, etc. capture_game_view— verify visually
macOS setup: System Settings → Privacy & Security → Screen Recording → enable Cursor (and Terminal if testing from CLI). Without this, tools fall back to text-only context.
Set OPENRCT2_VISION_MAX_WIDTH=1280 (default) to limit screenshot size.
Agent skill
See skills/openrct2-park-director/SKILL.md for orchestration workflows (perimeter coasters, cute themes, staff patrol).
Reload ride-builder after updates: ./scripts/build-plugins.sh then restart OpenRCT2 (plugin loads at launch).
Ride operations (P6)
| Tool | Description |
|---|---|
set_ride_inspection_interval_tool |
Inspection every 10–120 minutes |
set_ride_mode_tool / set_num_trains_tool / set_cars_per_train_tool |
Throughput tuning |
optimize_ride_throughput_tool |
Composite fix for struggling rides (dry_run supported) |
demolish_ride_tool |
Demolish any ride (confirm_destructive=true) |
manage_paths |
Now supports remove_tile and remove_line |
Land, finance, guests (P7–P9)
| Tool | Description |
|---|---|
find_open_land_tool / terraform_region_tool / buy_land_tool / clear_area_tool |
Land survey and terraforming |
get_finance_summary_tool / optimize_park_pricing_tool / start_marketing_campaign_tool |
Business loop |
fund_research_tool / scenario_progress_tool |
Research and objectives |
get_complaint_hotspots_tool / guest_flow_summary_tool |
Guest intelligence |
Advanced building and safety (P10–P11)
| Tool | Description |
|---|---|
coaster_build_along_path_tool / import_track_design_tool |
Path-following coasters, clipboard designs |
place_ride_at_best_tile_tool / extend_queue_tool |
Smart placement |
pre_change_snapshot_tool / get_action_log_tool |
Safety checkpoints and action journal |
| Destructive tools | Require confirm_destructive=true; composites support dry_run=true |
Performance
Large parks (1000+ guests, many rides) make bulk bridge queries slow or prone to timeout.
The MCP server avoids the worst patterns:
| Slow (avoid) | Fast (used instead) |
|---|---|
rides (all at once) |
listAllRides + rides?id=N per ride |
guests (all peeps) |
park.guests count + get_guest(id) |
game.state.park() for overview |
Scalar park.name, park.rating, park.cash, … |
| Tile-by-tile map scans | Not exposed via MCP (too many round-trips) |
| Footpath scan (fill benches/bins) | get_elements_by_type("footpath") bulk first; tile chunk fallback |
list_guests defaults to guest count only. Pass full_scan=true only if you accept a long wait.
Reload the MCP server after pulling performance updates.
Configuration
Environment variables (set in .cursor/mcp.json):
| Variable | Default | Purpose |
|---|---|---|
OPENRCT2_BRIDGE_PORT |
20020 |
Bridge plugin TCP port |
OPENRCT2_RIDE_BUILDER_PORT |
20021 |
Ride-builder TCP port |
OPENRCT2_BRIDGE_TIMEOUT |
30 |
Per-request bridge timeout (seconds) |
OPENRCT2_RIDE_BUILDER_TIMEOUT |
15 |
Per-request ride-builder timeout (seconds) |
OpenRCT2 config.ini recommendations:
- macOS:
~/Library/Application Support/OpenRCT2/config.ini - Linux:
~/.config/OpenRCT2/config.ini
Override the user data directory with OPENRCT2_USER_PATH (used by install-bridge.sh).
[plugin]
enable_hot_reloading = true
install-bridge.sh enables hot reloading automatically.
Contributing
./scripts/install-bridge.sh— Python venv, bridge plugin, hot reload./scripts/build-plugins.sh— ride-builder pluginpytest mcp-server/tests -q— offline unit tests (no OpenRCT2 required).venv/bin/python scripts/verify-e2e.py— optional smoke test with OpenRCT2 running
Development
# Run MCP server manually
export PYTHONPATH=mcp-server
.venv/bin/python -m openrct2_mcp
# Rebuild ride-builder after edits
./scripts/build-plugins.sh
# Reload plugin in-game (hot reload) or restart OpenRCT2
Safety notes
- Building tools pause the game before mutations.
build_in_pause_modecheat is enabled on connect for automation.- Coaster tools require the ride-builder plugin; park tools require openrct2-bridge.
License
MIT. See NOTICE for third-party attribution. Ride-builder plugin derived from openrct2-ridecreation-api.
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。