A Model Context Protocol server that lets an AI agent work on a real RPG Maker MV project on disk — database, maps, events, plugins, system — through 13 consolidated tools validated against the actual engine, so what comes out is coherent and playable.

It does three things that are usually missing:

⚡ Quick start

1 — Add it to your MCP client. No clone needed; the package ships an executable.

{
  "mcpServers": {
    "rpgmaker-mv": {
      "command": "npx",
      "args": ["-y", "rpgmaker-mv-mcp"],
      "env": {
        "RPGMAKER_PROJECT_PATH": "C:/path/to/your/RPGMakerMV/project"
      }
    }
  }
}

# Claude Code, user scope
claude mcp add rpgmaker --scope user \
  --env RPGMAKER_PROJECT_PATH="C:/path/to/project" \
  -- npx -y rpgmaker-mv-mcp

# From source
git clone https://github.com/DiegoLopez0208/RpgMakerMVUltimate-MCP
cd RpgMakerMVUltimate-MCP
npm install && npm run build
RPGMAKER_PROJECT_PATH=/path/to/your/project npm start

2 — Point it at a project. RPGMAKER_PROJECT_PATH is the folder containing data/, js/ and index.html. The server starts without it; call set_project_path at runtime instead if you prefer.

3 — Let the agent look around first.

get_project_context { detail: "full" }        → what exists, with real IDs
analyze_project     { view: "overview" }      → health, counts, unreachable maps

Works with Claude Desktop, Claude Code, opencode, and any MCP-compatible client.

🧭 What it does

flowchart LR
    A["🤖 Agent"] -->|"generate_map · manage_map_event"| B["📁 Project on disk"]
    B -->|"validate · balance · metrics"| A
    B -->|"playtest"| C["🎮 Running game"]
    C -->|"exceptions · position · screenshots"| A
    A -->|"reload_map"| C

The bottom half of that loop is what the bridge adds. Before it, the agent wrote files and hoped.

🗺️ Map generation

Two paths, both behind generate_map. Pick by whether your project uses RTP art.

The knowledge-driven path

{ "mode": "procedural", "theme": "town", "name": "Riverbend", "width": 40, "height": 30 }

Themes with a matching template — town, village, dungeon, interior, castle, world and more — clone a real map instead of painting tile noise. Themes without one (beach, swamp, desert…) fall back to Perlin terrain, BSP dungeons and cellular caves. Combat themes auto-wire random encounters from your existing troops; town and village auto-create enterable house interiors with two-way warps.

Themes · forest town village castle dungeon cave beach desert swamp ruins interior snow harbor volcano sewer fortress magic_forest magic_interior space_interior space_exterior world

Other modes: blank (empty canvas), themed (simple layout), template (one specific bundled map), batch (many at once), duplicate (copy an existing map).

The tileset-independent path

The bundled templates are raw MV map JSON, so their tile IDs only mean anything on RTP sheets. Change the tileset and the map turns to noise. semantic keeps the layout abstract until the last moment:

manage_system { action: "mine_templates" }               # learn from THIS project
generate_map  { mode: "semantic", tilesetId: 5, rooms: 6, seed: 42 }

• Mining reads every map you already made and derives semantic layouts (ground / wall / water / prop / door, multi-tile props kept whole), a tileset profile naming the concrete tile your project uses for each role, and token adjacency counts. Nothing in the project is modified — everything lands in .mcp-cache/.
• Generating builds the mission first — entrance → key → locked door → treasure → boss → exit, plus side rooms — as a graph whose edges are the only ways through, then paints it. Because the lock is an edge and the key sits on the entrance side of it, the map is solvable by construction. Autotile shapes are recomputed at the end from the finished neighbourhood, never guessed cell by cell.
• The result includes markers naming the cell of every mission role, which is where to put events with manage_map_event.
• Pass a mined templateId (e.g. "mined-3") to re-materialise one of your own maps onto a different tileset.

🔌 The live bridge

playtest on its own is fire-and-forget: the game opens and nothing comes back. The bridge closes the loop.

manage_system { action: "install_bridge_plugin" }   # once per project
manage_system { action: "bridge_start" }            # opens ws://127.0.0.1:32123
manage_system { action: "playtest" }                # the game connects on its own
manage_system { action: "bridge_telemetry", types: ["exception", "log"] }

sequenceDiagram
    participant A as 🤖 Agent
    participant S as 🖥️ MCP server
    participant G as 🎮 Game (nwjs)
    A->>S: edit_map
    S->>S: atomic write to Map002.json
    A->>S: bridge_command reload_map
    S->>G: reload_map
    G->>G: reserveTransfer + _needsMapReload
    G-->>S: reload_complete
    A->>S: take_screenshot
    S->>G: capture_screenshot
    G-->>S: PNG in base64
    S-->>A: path for analyze_image

• 📡 Telemetry — exceptions with stack traces, console.error/warn, scene changes, player position, which event command is executing (so a hung event can be pinpointed), FPS and heap. Frames are consumed as you read them unless you pass peek.
• ♻️ Hot reload — reload_map re-reads the current MapXXX.json and rebuilds the scene without losing party state: it reserves a transfer to the player's own position with _needsMapReload, the engine's own reload seam, rather than rebuilding Spriteset_Map by hand. reload_database re-reads one data file; System.json and Tilesets.json need a fresh playtest and are refused with an explanation.
• 📸 Screenshots — take_screenshot { name: "collision-proof" } captures the live playtest through the MCP plugin, saves a timestamped PNG under .mcp-cache/screenshots/, and returns its path for inspection or QA evidence. No shell screenshot command is involved. manage_system { action: "bridge_screenshot" } remains as a compatibility alias.

🔒 Security

The plugin returns before anything else runs unless the game is under NW.js and was launched with a test argument. A deployed build a player double-clicks never reaches the socket code, or even require('fs').

It checks every argument rather than only argv[0] the way Utils.isOptionValid does, because playtest passes the project path first. So a deployed build deliberately launched with a literal test argument would get past the guard — and then find no handshake file, and never connect.

The server binds 127.0.0.1 only, refuses any upgrade carrying a browser Origin (cross-site WebSocket hijacking), and requires the session token from .mcp-bridge.json — compared in constant time — within 5 seconds or the connection is dropped.

The command surface is a fixed allowlist with no eval primitive.

🔍 Project intelligence

analyze_project is read-only and fully offline. It models the whole project once, so an agent can reason about a game it did not build.

Offline map inspection

• query_map { view: "ascii", mapId } — render a map as a character grid with event markers. The cheapest way to see a layout and pick coordinates.
• query_map { view: "validate", mapId } — lint one map for invalid tile IDs, broken transfers and missing event terminators.

🧰 The 14 tools

🛡️ Write safety

• Atomic. Every write goes to a temp file and is renamed over the target, so an interrupted call can never leave half-written JSON.
• Backed up. Rotated timestamped copies under .mcp-backups/ (last N, RPGMV_BACKUP_KEEP, default 10).
• Previewable. Pass dryRun: true to any mutating tool to see exactly what it would write, without touching disk.

⚠️ Close the RPG Maker editor while an agent is working. The editor holds the project in memory and will overwrite changes when it saves.

⚙️ Configuration

# OpenAI
VISION_API_URL=https://api.openai.com VISION_API_KEY=sk-... VISION_MODEL=gpt-4o npm start
# Ollama (local, no key)
VISION_API_URL=http://localhost:11434 VISION_MODEL=llava npm start

🎓 Agent Skill

A portable Agent Skill teaches any model the crash-free workflow — build maps with generate_map, add content with manage_map_event presets, never hand-paint tiles or guess IDs. It lives at skill/rpgmaker-mv-mcp/SKILL.md.

# Claude Code / Claude.ai
npx degit DiegoLopez0208/RpgMakerMVUltimate-MCP/skill/rpgmaker-mv-mcp ~/.claude/skills/rpgmaker-mv-mcp
# opencode
npx degit DiegoLopez0208/RpgMakerMVUltimate-MCP/skill/rpgmaker-mv-mcp ~/.opencode/skills/rpgmaker-mv-mcp

Also listed in awesome-claude-skills.

📚 Knowledge base

🚧 Known limitations & roadmap

• Decoration semantics are best-effort in the RTP-template path; rare multi-tile objects may land as single tiles. The mined path keeps multi-tile props whole.
• Town and dungeon layouts keep improving — planned: a central plaza or well as a landmark, houses in rows facing roads, fences and yards, richer road networks, more room variety.
• mode: "semantic" currently generates dungeon-shaped missions. Town and open-world mission grammars are next, as is using the mined adjacency counts to decorate rather than only to describe.
• balance compares like with like inside a category, so a boss will legitimately look like an outlier next to random encounters. Read the flag, not the verdict.
• The bridge is Windows/nwjs playtest only and needs its plugin installed in the project.
• Vision AI requires your own endpoint.

🛠️ Development

npm install
npm run build      # tsc compile (+ copies knowledge/ into dist/)
npm test           # vitest
npm run typecheck
npm run dev        # tsx watch mode

💬 Feedback

Actively developed, and feedback is very welcome — bug reports, weird maps, missing tools, ideas. Open a GitHub Issue with what you asked the agent to do and what you got; an exported map JSON or a screenshot helps a lot.