A first-class meta-skill for creating, validating, and reviewing DCC-MCP skill
packages. It bundles scaffold/validation tools together with agent-facing
authoring guidance for SKILL.md, tools.yaml, scripts, groups, prompts, and
progressive-loading taxonomy.
Use dcc-mcp-creator when the task is to create a full adapter repository for
a host such as Nuke, Blender, 3ds Max, Unreal, ZBrush, Houdini, or Maya. Use
this skill when the task is to create or improve the skill packages loaded by
those adapters.
This skill ships with dcc-mcp-core. Add it to your skill path:
# Linux/macOS
export DCC_MCP_SKILL_PATHS="${DCC_MCP_SKILL_PATHS}:$(python -c 'import dcc_mcp_core; print(dcc_mcp_core.__file__)')/../skills"
# Windows
set DCC_MCP_SKILL_PATHS=%DCC_MCP_SKILL_PATHS%;C:\path\to\dcc-mcp-core\skills
Or reference it directly when starting your MCP server:
from dcc_mcp_core import create_skill_server, McpHttpConfig
server = create_skill_server(
"maya",
McpHttpConfig(port=8765),
extra_paths=["/path/to/dcc-mcp-core/skills"],
)
# Call the loaded MCP tool:
# dcc_mcp_skills_creator__create_skill(
# name="maya-rigging",
# parent_dir="/path/to/skills/dir",
# dcc="maya",
# tool_name="create_locator",
# affinity="main",
# )
from dcc_mcp_core import validate_skill
report = validate_skill("/path/to/my-skill")
if report.has_errors:
for issue in report.issues:
print(f"[{issue.severity}] {issue.category}: {issue.message}")
else:
print("Skill is valid!")
# Call the loaded MCP tool:
# dcc_mcp_skills_creator__skill_template()
my-skill/
|-- SKILL.md # Required: metadata frontmatter + instructions
|-- tools.yaml # Required when metadata.dcc-mcp.tools points here
|-- scripts/ # Optional: tool implementation scripts
| `-- create_locator.py
`-- references/ # Optional: recipes, examples, and long-form docs
|-- RECIPES.md
`-- NOTES.md
Generated tools.yaml entries follow the modern contract:
__ metadata.dcc-mcp.version in SKILL.md; a top-level version key is rejected by the strict loader.
metadata.dcc-mcp.depends as skill names,not repo names or prose-only instructions. Use it when one skill must be
discovered or loaded before another, for example depends: ["qt-ui-inspector"].
input_schema and output_schema are declared explicitly.input_schema shapes simple: prefer a top-level object with properties, required, primitive type, bounds, and descriptions. Put
mutually exclusive forms, conditional requirements, and cross-field rules in
the tool script or handler validation instead of anyOf, oneOf, allOf,
not, if/then/else, or dependent-schema keywords.
execution is sync or async; use async for deferred/long-running work.affinity is explicit. Use main for host API or scene mutation work and any for pure work.enforce_thread_affinity: true is emitted so adapter dispatch stays honest.annotations use MCP hints: read-only, destructive, idempotent, open-world, and deferred.call_examples: optional list of ready-to-copy argument payloads. Each entry has arguments (JSON object matching input_schema.properties) and an optional note. Surfaced in describe responses at metadata.dcc.call_examples so agents can construct correct arguments on the first attempt.dcc_mcp_core.skills_helper first: JSON/YAML codecs, bounded HTTP helpers, safe file/path helpers, validation, cancellation checks, and result helpers.metadata.dcc-mcp.depends for prerequisite skills, then declare execution, affinity, timeout_hint_secs, schemas, annotations, and failure recovery chains in tools.yaml. For high-frequency tools, add call_examples so agents can copy argument payloads without trial-and-error.references/.validate_skill_dir or dcc_mcp_core.validate_skill() before loading it in an adapter.When reviewing existing skills, reject top-level DCC-MCP extension keys such
as dcc, version, tags, tools, groups, depends, search-hint,
runtimes, prompts, and resources. Move them under
metadata.dcc-mcp.*; for version metadata, use
metadata.dcc-mcp.version: "1.0.0". Validate the installable skill directory
that contains the SKILL.md loaded by adapters, not only mirrored repository
docs or marketplace metadata.
Read AUTHORING_WORKFLOW.md and
DCC_TOOL_CONTRACTS.md before changing a
production skill package.
Gateway search treats tags as a narrowing filter. Use a small shared vocabulary
so pipeline, production-tracking, and documentation connectors rank and filter
consistently across hosts. When authoring SKILL.md frontmatter, include the
appropriate tags under metadata.dcc-mcp.tags:
| Tag | Use for |
|---|---|
| ----- | --------- |
pipeline | Studio pipeline systems, publish/intake/review automation, and production data hand-offs. |
production-tracking | Shot/asset/task/status tracking systems regardless of vendor. |
shotgrid | Autodesk Flow Production Tracking / ShotGrid-specific tools. |
ftrack | ftrack-specific tools. |
docs | Documentation, product help, reference lookup, and guide resources. |
read-only | Discovery/read operations. Also set MCP readOnlyHint (annotations.read_only_hint: true in tools.yaml); the tag is for search, not policy. |
destructive | Mutating or irreversible operations. Also set MCP destructiveHint (annotations.destructive_hint: true in tools.yaml); the tag is for search, not policy. |
Filter semantics:
dcc_type (singular) + dcc_types[] — OR: a result matching any listed DCC family passes. Include dcc_type: "maya" with dcc_types: ["blender"]
to match records from either host in one request.
tags[] — AND: a result must carry every listed tag. Use pipeline + production-tracking to narrow to records that carry both.
tags_any[] — OR: a result carrying any listed tag passes. Combines with the AND filter above: tags: ["pipeline"] + tags_any: ["read-only", "docs"]
returns pipeline records that are read-only OR documentation.
Vendor tags can be added when they sharpen routing without replacing the
canonical tags. For example, Autodesk Product Help should use docs,
read-only, and the vendor tag autodesk. Do not add docs to a
production-tracking search unless the user explicitly asks for help or reference
material.
Skill SKILL.md example (frontmatter excerpt):
metadata:
dcc-mcp:
dcc: shotgrid
layer: domain
tags: [pipeline, production-tracking, shotgrid]
search-hint: "ShotGrid task status, find shots, update task assignments"
tools: tools.yaml
# Read-only docs connector (SKILL.md excerpt)
metadata:
dcc-mcp:
dcc: autodesk-help
layer: infrastructure
tags: [docs, autodesk, read-only, infrastructure]
search-hint: "Autodesk Product Help, Maya help, 3ds Max help, API reference"
tools: tools.yaml
Individual read tools should also carry read-only in their tool-level tags;
mutating publish/update tools should carry destructive when applicable.
The validator checks:
name, descriptionsource_file references exist in scripts/metadata.dcc-mcp.tools/groups/prompts references existmetadata.dcc-mcp.depends consistencymetadata.dcc-mcp.* and point to sibling filesmetadata.dcc-mcp.version is accepted and projected to SkillMetadata.version; top-level version fails with an actionable
migration hint
validate_skill_dir emits skill-helper-adoption warnings when scripts import avoidable dependencies covered by dcc_mcp_core.skills_helper, such as requests, httpx, PyYAML, or local JSON/HTTP/file/path helper modules共 25 个版本