cocos-mcp-v3-ui

cocos-mcp-v3-ui

An MCP server that lets Claude Code drive the Cocos Creator 3.8.x editor to build UI, enabling translation of Figma designs into Cocos scenes by creating nodes, components, sprites, and prefabs.

Category
Visit Server

README

cocos-mcp-v3-ui

MCP server (UI-design tools) for Cocos Creator 3.8.x, built as an editor package. It lets an MCP client (Claude Code) drive the editor to build UI — ideal for translating a Figma design into a Cocos scene: read the design with the Figma MCP, then create nodes/components/sprites/prefabs here.

Ported (not mechanically) from a 2.x cocos-mcp-v2-ui. 3.x differs fundamentally: the scene bridge is the Promise-based Editor.Message.request('scene','execute-scene-script'), node size/anchor/opacity/color live on components (cc.UITransform / cc.UIOpacity / the render component — not on the node as in 2.x), and the package is plain CommonJS.

Architecture

Claude Code ──stdio(MCP)──▶ bridge/  ──HTTP /rpc──▶ editor package (main proc)
                          (system Node,            core/server + registry + tools
                           @modelcontextprotocol)        │
                                                          ▼  Editor.Message.request
                                                             ('scene','execute-scene-script')
                                                   scene.js (scene proc, cc.* live)
  • Editor package (main.js, core/, tools/, scene.js, scene/) runs inside Cocos Creator. main.js starts an HTTP server (core/server.js) exposing GET /health, POST /rpc {tool,args}, GET /tools, and a stateless POST /mcp. The server is loopback-only and rejects browser-originated requests (no Origin header, loopback Host only) so a web page cannot drive the editor (see isLocalRequest).
  • Bridge (bridge/) is a separate Node process Claude spawns. It speaks the MCP protocol over stdio (via the official SDK) and proxies tools/list+tools/call to the editor's HTTP /tools+/rpc. Keeping MCP-spec compliance out of the editor process.
  • Single-responsibility / DRY: tool modules are thin (schema + delegate). Everything crossing into the scene goes through one core/scene-bridge.js#callScene; one core/response.js shape; scene-side has one resolveNode, one serialize, one applyTransform. Each mutating tool records an editor snapshot (undo + dirty flag).

Install

  1. Symlink (or copy) this folder into the project's extensions/ (or packages/):
    ln -s /Users/nguyennt/Documents/MCP/cocos-mcp-v3-ui \
          <your-cocos-3.x-project>/extensions/cocos-mcp-v3-ui
    
  2. Install the bridge deps once: cd bridge && npm install.
  3. Open the project in Cocos Creator 3.8.x. The package auto-starts the server on port 8585 (autoStart: true). Panel: menu Extension → Cocos MCP v3 UI → Open Panel.

Connect to Claude Code

claude mcp add cocos-ui --transport stdio -- \
  node /Users/nguyennt/Documents/MCP/cocos-mcp-v3-ui/bridge/bin/cocos-ui-bridge.js --port 8585

Verify: /mcp in Claude lists cocos-ui connected and the 36 tools. The editor must be running with a scene open for scene tools to work (otherwise they return No active scene).

Tools (36)

Category Tools
ping (1) ping
scene (5) get_current_scene, get_scene_hierarchy, open_scene, save_scene, capture_node
node (10) create_node, get_node_info, find_nodes, set_node_transform, set_node_property, set_node_properties_batch, create_node_tree, set_sibling_index, delete_node, duplicate_node
component (13) add_component, set_component_property, get_components, remove_component, set_component_reference, add_click_event, add_components_batch, set_component_references_batch, add_click_events_batch, set_label_style, apply_widget, draw_rounded_rect, apply_gradient
asset (2) import_asset, query_asset_uuid
prefab (2) create_prefab, instantiate_prefab
referenceImage (3) add_reference_image, set_reference_image_data, remove_reference_image

MCP tool name = <category>_<method> (e.g. node_create_node, referenceImage_add_reference_image). A category must not contain an underscore — dispatch splits on the FIRST underscore.

Highlights (added per build-experience feedback)

  • color is forgiving: set_component_property/set_node_property/tree specs accept {r,g,b,a}, [r,g,b,a], or a hex string ("#161718").
  • create_node_tree: build an entire node + component + property subtree in ONE call (returns uuids).
  • draw_rounded_rect: native rounded backgrounds/borders via cc.Graphics — no texture hacks.
  • apply_gradient: bakes a rounded gradient PNG and binds it as a spriteFrame — real gradient buttons/badges (cc.Graphics fill is solid-only).
  • set_component_reference + add_click_event: wire ScrollView.content, Widget.target, button handlers, etc.
  • Enum names: enum props accept names (type:"VERTICAL", sizeMode:"CUSTOM") as well as numbers.
  • import_asset auto-creates the destination folder, auto-rasterizes SVG-content files, and fails loudly instead of silently importing nothing. spriteFrame assignment preserves node size.

Figma → Cocos workflow

  1. Read the design with the Figma MCP (positions, sizes, colors, text).
  2. scene_get_scene_hierarchy → find the Canvas uuid.
  3. (optional) export a screenshot from Figma → asset_import_assetreferenceImage_add_reference_image { spriteFrameUuid } to trace it.
  4. node_create_node per element → component_add_component (cc.Sprite/cc.Label/ cc.Button/cc.Widget/cc.Layout) → component_set_component_property (string, fontSize, color, spriteFrame) → node_set_node_transform to place/size.
  5. For images: asset_import_asset returns spriteFrameUuid — pass that to set_component_property property:"spriteFrame" (it persists; the texture uuid does not).
  6. Reusable bits: prefab_create_prefabprefab_instantiate_prefab.
  7. scene_save_scene.

Notes & limitations (3.x specifics)

  • capture_node (edit-mode) — 3.x has no immediate-mode camera.render(node), so we render through a temporary off-screen cc.Camera + cc.RenderTexture framed on the node's world bounds, force a frame, then readPixels() (falls back to borrowing the scene camera). The scene side returns base64 RGBA (bottom-up); the main side flips + PNG-encodes to disk and returns the path. Works in edit mode at the node's own aspect; maxSize caps the longest side (default 720). You can also verify by DATA with get_scene_hierarchy / get_components.
  • One render component per node (3.8.x)cc.Sprite, cc.Graphics, cc.Label, cc.RichText all derive from cc.UIRenderer and remain mutually exclusive on a node (verified: adding cc.Sprite next to cc.Graphics throws a conflict error). So a node can't have both draw_rounded_rect (Graphics) and apply_gradient/spriteFrame (Sprite) — apply_gradient auto-routes to a child gradient_bg node when the target already has a render component, and add_component surfaces the engine's conflict message.
  • Reference image is simulated — Cocos has a native reference-image gizmo, but to keep the workflow we add a low-opacity Sprite node __ReferenceImage__ behind the UI. Remove it before shipping (remove_reference_image); it is a real scene node.
  • Editor Hierarchy/Inspector refresh on the next reselect/reload after raw-runtime mutations. Each mutating tool records a snapshot (so undo works and the scene is marked dirty); the live scene/canvas updates immediately.
  • No main-process hot-reload: editing main.js/core//tools//scene/ during development requires a full editor restart (Developer → Reload Extensions/Editor). At runtime this is irrelevant — the server starts on project open.
  • spriteFrame: importing a PNG yields a texture and a sprite-frame sub-asset. Bind the sub-asset uuid (returned as spriteFrameUuid). In 3.x these sub-assets have stable uuid suffixes (<uuid>@f9941 spriteFrame, <uuid>@6c48a texture); assetdb.importFiles verifies them via asset-db:query-url before returning.
  • Verified editor messages (3.8.x): scene execute-scene-script / save-scene / open-scene / snapshot; asset-db:query-uuid / query-url / import-asset / create-asset / refresh-asset. Prefab create uses the scene create-prefab message where available, with a manual drag-to-Assets fallback.

Security

The HTTP server binds 127.0.0.1 and only serves loopback requests that carry no Origin header (i.e. the Node bridge, not a browser). This prevents a malicious web page from POSTing to http://127.0.0.1:8585/rpc to drive the editor or write files, and the loopback Host check blocks DNS-rebinding. Do not expose the port beyond localhost.

Tests

  • Unit (pure logic — colors, enums, PNG encoder, gradient, SVG sniff; no editor needed):
    node test/unit.test.js
    
  • Integration (error/edge paths for the scene & component tools; needs the editor running with a scene open — creates a temp container node and deletes it afterwards):
    node test/integration.test.js --port 8585
    
  • See test/manual-checklist.md for the per-tool manual verification steps.

Recommended Servers

playwright-mcp

playwright-mcp

A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.

Official
Featured
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.

Official
Featured
Local
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.

Official
Featured
Local
TypeScript
VeyraX MCP

VeyraX MCP

Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.

Official
Featured
Local
graphlit-mcp-server

graphlit-mcp-server

The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.

Official
Featured
TypeScript
Kagi MCP Server

Kagi MCP Server

An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.

Official
Featured
Python
E2B

E2B

Using MCP to run code via e2b.

Official
Featured
Neon Database

Neon Database

MCP server for interacting with Neon Management API and databases

Official
Featured
Exa Search

Exa Search

A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.

Official
Featured
Qdrant Server

Qdrant Server

This repository is an example of how to create a MCP server for Qdrant, a vector search engine.

Official
Featured