██████╗ ██╗      █████╗ ██╗   ██╗ ██████╗ █████╗ ███╗   ██╗██╗   ██╗ █████╗ ███████╗
██╔══██╗██║     ██╔══██╗╚██╗ ██╔╝██╔════╝██╔══██╗████╗  ██║██║   ██║██╔══██╗██╔════╝
██████╔╝██║     ███████║ ╚████╔╝ ██║     ███████║██╔██╗ ██║██║   ██║███████║███████╗
██╔═══╝ ██║     ██╔══██║  ╚██╔╝  ██║     ██╔══██║██║╚██╗██║╚██╗ ██╔╝██╔══██║╚════██║
██║     ███████╗██║  ██║   ██║   ╚██████╗██║  ██║██║ ╚████║ ╚████╔╝ ██║  ██║███████║
╚═╝     ╚══════╝╚═╝  ╚═╝   ╚═╝    ╚═════╝╚═╝  ╚═╝╚═╝  ╚═══╝  ╚═══╝  ╚═╝  ╚═╝╚══════╝

███╗   ███╗ ██████╗██████╗        ███████╗███████╗██████╗ ██╗   ██╗███████╗██████╗ 
████╗ ████║██╔════╝██╔══██╗       ██╔════╝██╔════╝██╔══██╗██║   ██║██╔════╝██╔══██╗
██╔████╔██║██║     ██████╔╝       ███████╗█████╗  ██████╔╝██║   ██║█████╗  ██████╔╝
██║╚██╔╝██║██║     ██╔═══╝        ╚════██║██╔══╝  ██╔══██╗╚██╗ ██╔╝██╔══╝  ██╔══██╗
██║ ╚═╝ ██║╚██████╗██║            ███████║███████╗██║  ██║ ╚████╔╝ ███████╗██║  ██║
╚═╝     ╚═╝ ╚═════╝╚═╝            ╚══════╝╚══════╝╚═╝  ╚═╝  ╚═══╝  ╚══════╝╚═╝  ╚═╝

An MCP Server for automating the PlayCanvas Editor using an LLM.

Important

At the moment, the MCP Server needs to be driven by Anthropic's Claude. Our experience shows that the free tier for Claude does not deliver a big enough chat context to operate the MCP Server reliably. Therefore, we strongly recommend subscribing to a Pro Claude account.

Available Tools

• Entity
  • list_entities
  • resolve_entities
  • create_entities
  • delete_entities
  • duplicate_entities
  • modify_entities
  • reparent_entity
  • add_components
  • remove_components
  • add_script_component_script
  • attach_script
• Asset
  • list_assets
  • create_assets
  • delete_assets
  • instantiate_template_assets
  • set_script_text
  • script_parse
  • set_material_diffuse
  • set_material_properties
• Scene
  • query_scene_settings
  • modify_scene_settings
• Store
  • store_search
  • store_get
  • store_download
• Viewport
  • capture_viewport
  • focus_viewport
• Runtime (live Launch instance)
  • launch_start
  • launch_stop
  • capture_runtime
  • read_runtime_logs
  • inject_input

Runtime Tools

The runtime tools drive a real Launch instance (the editor's Launch button) so you can verify that a scene actually runs, not just how it looks at edit time:

• launch_start opens https://launch.playcanvas.com/<sceneId>?debug=true in a new window. The extension injects a content script there that connects back to the MCP server as the "runtime" peer. launch_start returns { url, sceneId, ready }.
• capture_runtime screenshots the running app (scripts/physics/animation active).
• read_runtime_logs returns the app's console output + uncaught exceptions/rejections (newest first, paginated; defaults to warnings + errors).
• inject_input dispatches keyboard / mouse / touch events to the running app (e.g. hold W for 500ms, click at a canvas coordinate, tap the screen), so you can drive end-to-end interactions and then verify with capture_runtime.
• launch_stop closes the launch window.

Notes:

• Allow pop-ups for the editor origin, otherwise launch_start cannot open the window.
• The launch page uses your existing PlayCanvas login session (same browser), so no extra auth step is needed.
• Reload the unpacked extension in chrome://extensions/ after updating it so the new launch content script + permissions take effect.

Response Format

Every tool returns a single, consistent envelope so agents can pattern-match on a stable shape:

{
  "data": <result> | null,   // business payload; an empty set is [], never an error
  "meta": {
    "tool": "entities:list",
    "status": "ok" | "error",
    "message": "...",          // present only on error; actionable, with a recovery hint
    // list tools also include pagination metadata:
    "total": 120, "count": 50, "hasMore": true, "nextCursor": "50"
  }
}

Notes for tool authors / agents:

• Errors never use a top-level error field; they set meta.status = "error" and put an actionable message in meta.message (the protocol-level isError flag is also set).
• Empty results (list_*, resolve_entities) are a successful empty list, not an error.
• Pagination: list_entities / list_assets accept limit (default 50) + offset. Page using meta.nextCursor (pass it back as offset) and stop when meta.hasMore is false.
• State snapshots: mutating tools (create_entities, modify_entities, add_components, reparent_entity, duplicate_entities, instantiate_template_assets, …) return the resulting entity/asset summaries — including a human-readable hierarchy path — so you rarely need a follow-up list_entities call.
• Annotations: read-only tools declare readOnlyHint, destructive tools (delete_*) declare destructiveHint, and store tools declare openWorldHint (they reach the network).
• Image tools (capture_viewport) return a protocol image block plus a parallel text block carrying the same meta.

Installation

Run npm install to install all dependencies.

Install Chrome Extension

1. Visit chrome://extensions/ and enable Developer mode
2. Click Load unpacked and select the extension folder
3. Load the PlayCanvas Editor. The extension should be loaded.

Run MCP Server

The MCP Server can be driven by Cursor or Claude Desktop.

Tip

We have found Claude Desktop to be generally more reliable.

Claude Desktop

1. Install Claude Desktop.
2. Go to Claude > Settings.
3. Select Developer and then Edit Config.
4. This will open claude_desktop_config.json, your MCP Config JSON file.

Cursor

1. Install Cursor.
2. Select File > Preferences > Cursor Settings.
3. Click + Add new global MCP server.
4. This will open mcp.json, your MCP Config JSON file.

Tip

Also in Cursor Settings, select Features and scroll to the Chat section. Activate Enable auto-run mode to allow the LLM to run MCP tools without requiring constant authorization. You do this at your own risk (but we prefer it)!

Important

In Cursor, ensure you have Agent selected. Ask and Edit modes will not recognize the MCP Server.

MCP Config JSON File

This is how your config should look:

Windows

{
  "mcpServers": {
    "playcanvas": {
      "command": "cmd",
      "args": [
        "/c",
        "npx",
        "tsx",
        "C:\\path\\to\\editor-mcp-server\\src\\server.ts"
      ],
      "env": {
        "PORT": "52000"
      }
    }
  }
}

macOS

{
  "mcpServers": {
    "playcanvas": {
      "command": "npx",
      "args": [
        "tsx",
        "/path/to/editor-mcp-server/src/server.ts"
      ],
      "env": {
        "PORT": "52000"
      }
    }
  }
}

Connecting the Editor to the MCP Server

The PlayCanvas Editor does not connect to the MCP Server automatically. To connect:

1. Activate a Chrome tab running the PlayCanvas Editor.
2. Select the Extensions icon to the right of the address bar.
3. Select PlayCanvas Editor MCP Extension to open the extension popup.
4. Select CONNECT (the port number should match what is set in your MCP Config JSON File).

Note

You can currently only connect one instance of the PlayCanvas Editor to the MCP Server at any one time.

You should now be able to issue commands in Claude Desktop or Cursor.