MCP + Skills Integration

中文版

Why We Combine MCP and Skills

The Problem with Stock MCP

Standard MCP tools/list returns every tool the server knows about — all at once.

With 3 DCC instances × 50 skills × 5 scripts = 750 tools in context. The AI's context window fills instantly, costs skyrocket, and reasoning quality degrades.

Stock MCP tools/list:
┌──────────────────────────────────────────────┐
│  maya_geo__create_sphere                      │
│  maya_geo__bevel                              │
│  maya_anim__set_keyframe  ... (x 250)         │
│  blender_sculpt__smooth                       │
│  blender_sculpt__grab     ... (x 250)         │
│  houdini_vex__compile     ... (x 250)         │
└──────────────────────────────────────────────┘
750 tools in context every time → expensive, slow

Our Solution: Session-Scoped Progressive Discovery

Each AI session is pinned to one DCC instance. tools/list is scoped to that instance.

Session A (Maya instance #1):
  tools/list → 100 Maya tools + 50 shared tools = 150 tools

Session B (Houdini instance #1):
  tools/list → 100 Houdini tools + 50 shared tools = 150 tools

71% reduction in context size, zero information loss

The Problem with CLI Tools

CLI tools are blind to DCC state:

• Can't see the active scene, selected objects, or viewport
• Require multiple roundtrips to gather context
• Return raw text requiring fragile parsing
• Have no visual feedback during execution

dcc-mcp-core lives inside the DCC and can access its full state directly.

Architecture

AI Agent (Claude, GPT, etc.)
    │
    │ tools/call {"name": "maya_geo__create_sphere", "radius": 2.0}
    │
    ▼
Gateway Server (dcc-mcp-server)          ← one per machine
    │
    │ Session A → Maya instance #1 (IPC)
    │ Session B → Houdini instance #1 (IPC)
    │
    ▼
DCC Bridge Plugin
    │
    │ Executes script, captures result
    │
    ▼
{success: true, message: "Sphere created", context: {name: "pSphere1"}}

Key Concepts

ServiceEntry — What Each DCC Registers

from dcc_mcp_core import create_skill_server, McpHttpConfig

# Gateway auto-registers the DCC instance
server = create_skill_server("maya", McpHttpConfig(port=8765))
handle = server.start()

The gateway reads this to know:

• Which instance is available
• What files are open
• Which instance to route the AI's request to

Session Isolation

# Session pinning is handled automatically by the gateway
# based on dcc_type, dcc_version, and scene metadata
# set on McpHttpConfig

Skills System

SKILL.md  (metadata)          scripts/
──────────────────────────────────────────────
name: maya-geometry           create_sphere.py
dcc: maya                     bevel.py
scope: repo                   export_fbx.bat
policy:                       ──────────────
  products: ["maya"]          ↓ registered as
  allow_implicit_invocation:  maya_geometry__create_sphere
    false                     maya_geometry__bevel
                              maya_geometry__export_fbx

Zero Python glue code. metadata + scripts = MCP tools.

Progressive Discovery

Tools are revealed gradually based on context:

Filter	Criteria	Result
DCC type	Session is pinned to Maya	Only Maya tools visible
Product	policy.products: ["maya"]	Houdini tools hidden
Scope	scope: system	Can't be overridden by repo skills
Implicit	allow_implicit_invocation: false	Requires explicit load_skill first

Quick Start

1. Install

pip install dcc-mcp-core

2. Create a skill

my-tools/
├── SKILL.md
└── scripts/
    └── create_sphere.py

SKILL.md:

---
name: my-tools
dcc: maya
scope: repo
policy:
  allow_implicit_invocation: true
  products: ["maya"]
---
# My Maya Tools
Custom geometry tools.

3. Start the server

import os
from dcc_mcp_core import create_skill_server, McpHttpConfig

os.environ["DCC_MCP_MAYA_SKILL_PATHS"] = "/path/to/my-tools"

server = create_skill_server("maya", McpHttpConfig(port=8765))
handle = server.start()
print(f"MCP server: {handle.mcp_url()}")
# AI clients connect to http://127.0.0.1:8765/mcp

4. Connect Claude Desktop

{
  "mcpServers": {
    "maya": {
      "url": "http://127.0.0.1:8765/mcp"
    }
  }
}

What Makes This Different

Feature	dcc-mcp-core	Generic MCP	CLI tools
DCC state awareness	✅ Scene, docs, objects	❌ None	❌ None
Context scoping	✅ Session-isolated	❌ Global	❌ N/A
Zero-code tools	✅ SKILL.md	❌ Full Python	✅ Scripts only
Multi-instance	✅ Gateway election	❌ Single endpoint	❌ No
Structured results	✅ Always	⚠️ Manual	❌ Text parsing