DCC-MCP Skills Creator

Create and validate DCC-MCP Skill packages

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.

Install and Route

Install the published @loonghao/dcc-mcp-skills-creator package, then start a new agent turn:

bash

openclaw skills install @loonghao/dcc-mcp-skills-creator
npx --yes clawhub@0.23.1 install @loonghao/dcc-mcp-skills-creator

Use dcc-mcp to operate an existing DCC and dcc-mcp-creator to build a complete adapter. A repository checkout may load this directory directly; DCC_MCP_SKILL_PATHS and extra_paths are runtime paths for DCC adapters, not installation instructions for an agent host.

CLI-First Control Path

Use the dcc-mcp skill and dcc-mcp-cli for skill discovery, loading, validation, and live calls whenever the agent can run shell commands. Start a live validation with dcc-mcp-cli list: if the process launches, the CLI is installed and the result checks the gateway plus DCC/MCP inventory. Diagnose a failed health/inventory result with dcc-mcp-cli doctor; do not reinstall the CLI, probe import dcc_mcp_core, or read server internals to infer readiness.

Only a shell-level command-not-found result means the CLI is missing. Ask for consent; after explicit approval, immediately run the verified dcc-mcp helper python scripts/check_cli.py --ensure-cli --pretty from that Skill's directory. It installs the official CLI and rechecks health/inventory in the same attempt, without a second confirmation. Keep it current with dcc-mcp-cli update check, then dcc-mcp-cli update apply; apply stages the next CLI launch and does not replace a running server binary.

Quick Start

Create a new skill

python

# 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",
# )

Validate an existing skill

bash

dcc-mcp-cli lint /path/to/my-skill

Get a SKILL.md template

python

# Call the loaded MCP tool:
# dcc_mcp_skills_creator__skill_template()

Skill Directory Structure

text

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

Current Tool Contract

Generated tools.yaml entries follow the modern contract:

• Local tool names are snake_case and client-safe. Do not use dotted names.
• Loaded tools are published as <skill-name>__<tool_name> when namespacing is needed.
• Skill package version metadata lives at metadata.dcc-mcp.version in SKILL.md; a top-level version key is rejected by the strict loader.
• Set metadata.dcc-mcp.dcc to the concrete host. Use dcc: any only when the same implementation is safe in every host; concrete-host tools override a same-named any tool during scoped lookup.
• Inter-skill dependencies live at 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.
• A zero-argument tool still declares a closed schema: {"type":"object","properties":{},"additionalProperties":false}.
• Runtime discovery never imports or executes tool scripts to infer missing schemas by default. Treat Python-derived schemas as an authoring-time helper: generate them before publishing, then commit the JSON Schema to tools.yaml.
• Keep MCP-facing 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. When a complex schema is unavoidable, discovery must route through describe before call.
• execution is sync or async; use async for deferred/long-running work.
• job_strategy is monolithic (default), chunked, or isolated. Agents use it to select a safe execution and recovery workflow.
• 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 explicitly declare boolean read_only_hint, destructive_hint, idempotent_hint, and open_world_hint. Missing safety fields force describe, even for a zero-argument tool; deferred_hint stays optional.
• Keep tool groups independently usable. A correlated load carrying target_tool_slug activates only that tool's group; do not rely on a sibling default-active group being activated with it.
• 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.

Long-Running Main-Affinity Tools

execution: async changes the job lifecycle; it does not make one monolithic host call interruptible. For long scene mutations:

yaml

execution: async
job_strategy: chunked
affinity: main
enforce_thread_affinity: true
annotations:
  deferred_hint: true

When the adapter supports HostUiDispatcherBase.submit_chunked_runner(), define bounded steps with the shared helper:

python

from dcc_mcp_core import chunked_job

@chunked_job(total=100)
def build_bake_steps():
    for frame in range(100):
        yield lambda frame=frame: bake_one_frame(frame)

Return the runner from the declarative entry point. HostExecutionBridge automatically submits it to the shared host pump and binds it to the outer JobManager cancellation probe. Do not create a skill-local timer, thread, pump, or second job registry. Keep each yielded callable bounded, return a string when a progress message is useful, and let cancellation become terminal only after a runner checkpoint. If the adapter does not expose the shared chunked path, document that the tool is monolithic and request an adapter/core integration instead of claiming mid-call interruption.

Use job_strategy: isolated when the typed tool launches a process- or service-owned operation and returns a durable job id immediately. Declare the poll and cancel tools in next-tools and in the result recovery context. Status must remain readable after a transport disconnect or adapter restart; state cancellation ownership honestly when it cannot be reconstructed.

Render and cook status tools should reuse the Core progress vocabulary: status, progress.current, progress.total, and progress.message. current and total are monotonic work-unit counts such as completed/total frames; clients derive the percentage and render one progress bar. Prefer the renderer or cook service's native counters. If files are the only source, keep that counting inside the typed status tool instead of making the agent run repeated directory scans.

During an active turn, agents should start once and use CLI --wait, REST job events, or the declared status tool. Do not create an OS or DCC-MCP scheduled workflow merely to poll one running operation. Only after the user explicitly requests cross-session monitoring may an agent create a one-shot follow-up that stores the existing job/operation id, performs read-only status checks, and self-stops at a terminal state; it must never relaunch the render or cook. Mirror this contract in agents/openai.yaml: tell the Agent to start once, follow typed progress to a terminal state, and query the same job id after a timeout instead of relaunching work.

For one indivisible DCC-native call, keep job_strategy: monolithic. Prefer execution: async so the initial transport returns a core job id, then poll the instance-routable jobs_get_status. A transport timeout is not completion or cancellation: rediscover the instance and query the job before retrying. The creator scaffold deliberately emits monolithic for async tools; change it only with the matching chunked runner or isolated status/cancel implementation.

Computer Use Fallback Contract

• Reuse the bundled ui-control skill instead of creating another screenshot, pointer, keyboard, or Windows SendInput tool set. Declare metadata.dcc-mcp.depends: ["ui-control"] only when it is a hard workflow dependency.
• Keep the visual loop as ui_control__snapshot -> ui_control__act -> ui_control__snapshot, and pass the latest snapshot_id unchanged. End every path with ui_control__stop_computer_use. Screenshot coordinates belong to that observation only.
• Preserve capture_provenance with saved evidence. Only backend=windows-ui-control-host plus pixels_captured=true proves native Windows screenshot capture; keep the logical UI Control session_id distinct from the gateway agent session used for stats attribution.
• When the exact HWND is minimized or hidden before the first snapshot, use only get_window_state followed by the necessary restore_window, show_window, and activate_window host actions, then take a fresh snapshot. Never substitute desktop enumeration or open-ended input.
• Stateful UI tools must declare requires_in_process: true independently of affinity; keep UI Control at affinity: any so it does not block the DCC UI thread while preserving one named-pipe client. On Windows, the isolated per-logon-session host owns observations, Esc interruption, confirmation, and the cross-adapter input owner; skill scripts must not instantiate an in-process ComputerUseSession fallback.
• Prefer a control_id and semantic UI Automation action. Use raw coordinates only when the UI does not expose a stable semantic control.
• For custom-drawn canvases, viewport manipulators, or face controls, use one drag path from the latest snapshot. keys may hold Ctrl, Shift, or Alt for pointer-modified drags; snapshot again immediately before deriving another path.
• Never set DCC_MCP_COMPUTER_USE_ALLOW_RAW_INPUT from a skill script. It is an operator-owned environment ceiling. Native input also requires the adapter/operator to bind its DCC with DCC_MCP_UI_CONTROL_UIA_PROCESS_ID or DCC_MCP_UI_CONTROL_UIA_WINDOW_HANDLE; a skill request may only narrow that trusted scope. Propagate user_interrupted immediately; do not retry the action or fall back to another input path after Esc interrupts a session.
• Never enter or retry another UI/input path after a policy, authorization, authentication, security, confirmation, desktop_unavailable, or user_interrupted result. Computer Use is a capability fallback, not a way around a control boundary.
• Keep mutating UI Control tools annotated as destructive. An optional consequence intent can only raise the native host's independent UIA/input classification. Never add a model-controlled confirmed or approved argument or treat an environment variable as per-action user approval.
• Generated record-replay Skills must stay local until reviewed. Compile structured calls to WorkflowSpec tool steps, compile semantic UI actions as fresh snapshot -> find -> one act -> verified wait/snapshot loops, and reject raw captured control ids or coordinates. Keep the demonstrated instance id as review provenance only. Never serialize approvals, grants, credentials, prompts, or secret-shaped fields. Visual fallback assets must be content-addressed, exact-window bounded, confidence gated, stable across multiple frames, and fail closed on geometry/DPI/topology drift.

Authoring Workflow

1. Decide whether the skill is infrastructure, domain, thin-harness, or example.

2. Give the skill a kebab-case name and each local tool a snake_case name.

3. Keep host API calls inside scripts, with lazy imports so discovery works without the host running.

4. Import same-directory helper modules directly; in-process runners expose the executing script's directory only for the call, so scripts must not mutate sys.path for sibling imports. In particular, do not repeat the legacy pattern shown in houdini#157:

   python

   script_dir = str(Path(__file__).resolve().parent)
   if script_dir not in sys.path:
       sys.path.insert(0, script_dir)
   

   That mutates process-global import state and leaks across skills. Script-directory lifetime is runtime ownership; use a direct sibling import and let the executor scope resolution to the current call.

5. Import dependency-light runtime helpers from dcc_mcp_core.skills_helper first: JSON/YAML codecs, bounded HTTP helpers, safe file/path helpers, validation, cancellation checks, and result helpers.

6. Declare metadata.dcc-mcp.depends for prerequisite skills, then declare execution, affinity, timeout_hint_secs, schemas, annotations, and failure recovery chains in tools.yaml. Do not rely on runtime Python introspection for missing schemas. For high-frequency tools, add call_examples so agents can copy argument payloads without trial-and-error.

7. Put long examples, recipes, and host-specific notes under references/.

8. Validate with validate_skill_dir or dcc_mcp_core.validate_skill() before loading it in an adapter. For discovery/load performance regressions, assert deterministic backend operation counts; use elapsed-time thresholds only as supplemental evidence.

9. If the desired behavior requires parsing core internals or adapter-private YAML at runtime, stop and request a core API instead.

Improve Skills From Completed Tasks

Use retained gateway evidence only after the user-visible task and its validation are complete. Keep one stable session_id in call metadata, then query the narrowest useful slice:

bash

dcc-mcp-cli stats --range 24h --dcc-type <dcc> --session-id <session-id>

Get the review_skill_improvement prompt from this skill and supply the stats JSON plus bounded task and validation summaries. Treat total_calls == 0 as missing evidence, not success. Never include hidden reasoning, raw prompts, credentials, or unredacted payloads.

Prefer no_change, then improving an existing skill, and create a new skill only for a repeated, reusable workflow that no current skill owns. Validate any accepted change with validate_skill_dir or dcc-mcp-cli lint before loading it. Statistics inform a proposal; they never authorize editing or publishing a skill without the task owner's requested scope.

For a failed task, first use the dcc-mcp recovery flow: retain the request_id, run doctor for runtime/readiness faults, query stats --status failure --session-id <session-id>, and record structured feedback through the CLI-discovered dcc_feedback__report tool. The public-safe /v1/debug/issue-reports/<request_id> payload is suitable for a reviewed issue; never publish ?mode=raw automatically.

Fix this Skill only when the evidence identifies its schema, script, description, next-tool, or workflow contract. Route adapter/runtime failures to dcc-mcp-creator and shared CLI/gateway/core failures to dcc-mcp-core. A one-off tool bug is not evidence for creating another Skill.

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-Facing Tag Taxonomy

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.

Python 3.7 Policy

All authored skills must declare compatibility: "Python 3.7+" in their frontmatter when they are installed into an LTS DCC host. This applies to every skill that is installed into a DCC host embedding Python 3.7 (Maya 2022, Blender 2.83, 3ds Max 2022, etc.). py37-lite is a supported fallback but does not replace the native Linux and Windows cp37 compatibility gates. See ADR 011 and compatibility/python.json for the deprecation and CI contract. In lite mode, create_skill_server() supports local metadata discovery (list_skills, search_skills, and get_skill) only. The Rust sidecar is dispatch-only, so gateway discovery and declarative load_skill execution require a native Python 3.7 wheel; lite activation fails explicitly.

For hermetic CI or tests, set DCC_MCP_DISABLE_DEFAULT_SKILL_PATHS=1 so an operator's local/platform defaults, marketplace installs, and Admin custom paths cannot alter discovery results. Explicit, bundled, and DCC_MCP_*_SKILL_PATHS paths remain active under this mode.

Skill SKILL.md example (frontmatter excerpt):

yaml

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

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.

Validation Rules

The validator checks:

• SKILL.md exists and is readable
• YAML frontmatter is well-formed
• Required fields: name, description
• Name format: kebab-case, <=64 chars, matches directory name
• Field lengths: description <=1024, compatibility <=500
• Tool declarations: non-empty names, no duplicates, snake_case client-safe format
• Script files: source_file references exist in scripts/
• Sidecar files: metadata.dcc-mcp.tools/groups/prompts references exist
• Dependencies: metadata.dcc-mcp.depends consistency
• Spec compliance: non-standard top-level keys are frontmatter errors; dcc-mcp-core extensions must live under metadata.dcc-mcp.* and point to sibling files
• Version metadata: metadata.dcc-mcp.version is accepted and projected to SkillMetadata.version; top-level version fails with an actionable migration hint
• Skill helper adoption: 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