The official distribution repository for DCC-MCP Agent Skills across Codex, Claude Code, CodeBuddy, WorkBuddy, OpenClaw, ClawHub, Gemini CLI, GitHub Copilot, Cursor, Windsurf, OpenCode, Cline, Roo Code, Kiro, Amp, and other Agent Skills-compatible hosts.
One canonical Skill suite lives under plugins/dcc-mcp/skills/. Vendor
manifests are thin adapters; they do not fork instructions or runtime behavior.
Released-product discovery is generated from
plugins/dcc-mcp/skills/dcc-mcp/references/PRODUCTS.json,
a provenance-bearing snapshot of the official dcc-mcp-cli dcc-types catalog
reconciled with the owning adapter repositories. It also carries a small
current-Core route set for adapters that have landed in Core before the next
CLI release. It carries one canonical identity and bounded aliases per route
instead of repeating hand-maintained product lists across manifests.
Install the default dcc-mcp Skill with one command. The installer detects
supported agents in the current workspace and lets you choose the target:
npx --yes skills@1.5.22 add dcc-mcp/dcc-mcp-agent-plugins --skill dcc-mcpThis covers Gemini CLI, GitHub Copilot, Cursor, Windsurf, OpenCode, Cline, Roo
Code, Kiro CLI, Amp, OpenHands, Continue, Replit, and other Agent
Skills-compatible hosts. Start a new agent session if the host discovers Skills
only at startup. Use --global for a user-level installation or
--agent <id> --yes for a non-interactive explicit target; see the installer's
supported-agent list.
After installation, prompts only need to describe the work:
Use the dcc-mcp Skill to <describe the DCC task>.
The Skill handles DCC discovery, typed tool selection, and approval boundaries.
Install dcc-mcp-creator or dcc-mcp-skills-creator only when developing an
adapter or authoring a reusable DCC-MCP Skill.
GitHub Release archives include a root plugin.json conforming to the
Agent Plugins 1.0.0 specification and the same
canonical skills/ directory. Compatible clients can load the archive as a
skills-only portable plugin; installation and distribution remain
client-defined by the standard.
Use a native marketplace when the host provides one. It installs the same canonical Skills.
codex plugin marketplace add dcc-mcp/dcc-mcp-agent-plugins
codex plugin add dcc-mcp@dcc-mcp/plugin marketplace add dcc-mcp/dcc-mcp-agent-plugins
/plugin install dcc-mcp@dcc-mcp
codebuddy plugin marketplace add dcc-mcp/dcc-mcp-agent-plugins
codebuddy plugin install dcc-mcp@dcc-mcpAdd this repository URL from Plugins > +, or upload a standalone Skill ZIP from the GitHub Release to Skills > Add skill > Upload skill.
openclaw skills install @loonghao/dcc-mcp
openclaw skills install @loonghao/dcc-mcp-skills-creator
openclaw skills install @loonghao/dcc-mcp-creator
openclaw skills install @loonghao/dcc-cuaDirect ClawHub installs are also supported:
npx --yes clawhub@0.23.1 install @loonghao/dcc-mcp
npx --yes clawhub@0.23.1 install @loonghao/dcc-mcp-skills-creator
npx --yes clawhub@0.23.1 install @loonghao/dcc-mcp-creator
npx --yes clawhub@0.23.1 install @loonghao/dcc-cuaUse the updater that owns the installation. OpenClaw records ClawHub installs, so workspace and shared Skills can be refreshed non-interactively:
openclaw skills update --all
openclaw skills update --all --globalA direct ClawHub install is tracked in the work directory where it was installed. Run the update from that same directory:
npx --yes clawhub@0.23.3 update --allSkills installed from GitHub with the universal Agent Skills CLI use its own lock file and updater:
npx --yes skills@1.5.23 update -p -y
npx --yes skills@1.5.23 update -g -yThese non-interactive commands can run from Task Scheduler, cron, or the host's
automation service. Keep --force out of unattended ClawHub updates so pinned
or locally modified Skills stay protected. The universal Agent Skills updater
reinstalls changed managed copies from their recorded source, so custom changes
belong in the source repository. An untracked directory copied into an agent's
Skill folder must be reinstalled once through one of these managers before
automatic updates can discover it.
The default Skill starts with a non-executing CLI availability check and then composes the existing host, adapter and update reports. Installed software, registered instances, readiness and a successful tool call stay separate. See First run for the report helper and safe legacy-CLI boundary, and Software requests for a reviewed, deduplicated issue draft. Neither flow installs software or publishes an issue without the user's approval.
After installing the Skill, start the DCC application and enable its DCC-MCP adapter. Then run:
dcc-mcp-cli listThe output must contain a running instance with the adapter's documented
dcc_type. If the command is missing, follow the adapter repository's setup
guide for the official CLI. If the command works but lists no matching
instance, keep the DCC application open and check that its extension, add-on,
or plugin is enabled. Restart the agent after installing a Skill or native
plugin when that agent discovers integrations only at startup.
Adapter repositories intentionally carry only a generated, short installation entry. This repository owns the canonical commands and troubleshooting links; see Maintaining adapter README quickstarts for the rollout and drift-check contract.
The four canonical Skills are mapped to Smithery's GitHub-backed Skills Registry. Search and install them with the Smithery CLI:
npm install -g smithery@latest
smithery skill search "dcc-mcp"
smithery skill add loonghao/dcc-mcp --agent claude-codeMaintainers can validate the mapping locally, then publish only with an
explicit flag and SMITHERY_API_KEY:
python scripts/smithery_sync.py
SMITHERY_API_KEY=... python scripts/smithery_sync.py --publishThe same canonical Skills ship as a package for hosts that vendor Skills
through node_modules:
npm install @dcc-mcp/skills
npx --yes skills@1.5.22 experimental_syncPulseMCP is an MCP Server directory, not a raw Agent Skills registry. This
repository therefore does not publish Skill ZIPs there. Submit a hosted or
local DCC-MCP Server from the owning adapter/core release through
pulsemcp.com/submit; keep Agent Skills on
Smithery, ClawHub, or an Agent Skills-compatible host.
Every release publishes a machine-readable catalog and GEO metadata to GitHub Pages, generated from the same manifests CI validates:
| File | Purpose |
|---|---|
catalog.json |
Skill metadata, install commands, and distribution channels |
llms.txt |
Curated entry point for AI crawlers and answer engines |
llms-full.txt |
Every SKILL.md concatenated in full |
sitemap.xml |
Per-Skill pages carrying schema.org/SoftwareApplication JSON-LD |
catalog.json also projects the complete released-product matrix and the
application UI route. Typed DCC-MCP tools remain first. DCC-CUA and
ui-control are searchable names for the same project-owned application UI
provider across DCC, browser, and non-DCC apps; explicit DCC-CUA requests never
fall back to generic Computer Use, Browser, or Chrome providers.
docs/DISTRIBUTION.md is generated from the same source and lists every
channel plus ready-to-paste entries for the directories that need a human
submission. CI fails when it drifts:
python scripts/build_geo_site.py # regenerate site/ and docs/DISTRIBUTION.md
python scripts/build_geo_site.py --check # fail when the committed doc is stale| Skill | Purpose |
|---|---|
dcc-mcp |
Operate live DCC applications through structured tools |
dcc-mcp-skills-creator |
Create and validate DCC-MCP Skills |
dcc-mcp-creator |
Create and modernize DCC-MCP adapters |
dcc-cua |
Route bounded DCC, browser, and non-DCC UI actions through project-owned DCC-CUA |
This repository is the sole source of truth for all four public Skills. It
owns their instruction bodies, helpers, tests, versions, product discovery
metadata, vendor manifests, archives, GitHub Releases, ClawHub, and Smithery.
Core owns runtime code, schemas, the CLI, and Core-only/runtime Skills; public
Skill bodies are no longer copied from Core. See
docs/adr/0003-agent-plugins-source-of-truth.md.
Conditional DCC-CUA routing is deliberately not declared as a hard depends
edge, so typed-tool tasks keep progressive loading and do not load the UI
provider unless UI behavior is actually required.
The Skill suite version is this repository's own release line, deliberately
decoupled from dcc-mcp-core release numbers: Skill content can be re-published
without a Core release, and Core can release without changing Skill content.
metadata.dcc-mcp.version in each SKILL.md follows this repository's suite
version.
Product discovery is a separate compatibility contract. It is regenerated from
PRODUCTS.json and can add a route from the current Core catalog (for example
the Office/PPT workflow) even while the released CLI snapshot remains at its
own version. OBS Studio and LiquiGen are part of the released CLI catalog. Run
sync_product_discovery.py --check --check-core-catalog to verify the generated
surfaces against the pinned released Core catalog.
When a user gives an absolute local application path, the default Skill stores
it in the per-user cache used by
scripts/app_path_cache.py.
Later requests reuse the path only after checking that it still exists and
asking the user to confirm startup. Missing paths produce a new-path prompt and
ask whether the host application needs installation. Free/open-source products
with host_install metadata show their official HTTPS source; downloading,
installing, and launching still require explicit user consent and remain
separate from the DCC-MCP adapter install hint. See
LOCAL_APP_PATH_CACHE.md.
# After editing a public Skill, bump the suite and regenerate distribution data.
python scripts/bump_version.py --patch
python scripts/sync_product_discovery.py
python scripts/build_geo_site.py
python scripts/sync_product_discovery.py --check --check-core-catalog --check-cliMerging any validated Skill version change publishes the four Skills to
ClawHub. Tagging v<version> creates the cross-channel GitHub Release,
Smithery, npm, and Pages release. A coordinated Core runtime change uses a
separate Core pull request; each repository reviews and validates its own
contract.
Publication gates check canonical manifest coverage and version increases against a trusted Git baseline. Its first-time activation and branch-protection requirements are documented separately from upload acceptance and public registry visibility.
python scripts/setup_released_core.py
python -m unittest discover -s tests -v
python scripts/sync_product_discovery.py --check
python scripts/validate_repository.py
npx --yes skills@1.5.22 add . --list
./scripts/build-packages.ps1
codex plugin marketplace add .
codex plugin add dcc-mcp@dcc-mcp
claude plugin validate . --strict
codebuddy plugin validate ./plugins/dcc-mcpTagging v<version> creates a GitHub Release with the plugin and standalone
Skill archives, then fans out to every automated channel: ClawHub, SkillHub, Smithery,
npm, and the GitHub Pages catalog. A final matrix job installs the released
Skills from the public repository on Linux, macOS, and Windows and asserts the
released version, so a broken public install fails the release.
| Secret | Channel |
|---|---|
CLAWHUB_TOKEN |
ClawHub |
SKILLHUB_KEY |
SkillHub (skillhub.cn) |
SMITHERY_API_KEY |
Smithery |
NPM_TOKEN |
npm (@dcc-mcp/skills, published with provenance) |
SkillHub uses the same four Skills and independent versions in
.github/clawhub-skills.json. The Sync SkillHub Skills workflow runs CLI
preflight on pull requests, supports manual dispatch (dry-run by default), and
publishes after the GitHub Release succeeds. Add SKILLHUB_KEY from
SkillHub API keys to repository secrets.
The CLI is pinned to 2026.8.5 with SHA-256 verification and self-upgrade disabled.
To upgrade it, update both the download URL and the verified archive digest.
scripts/stage_skillhub.py copies packable files into a fresh directory under
dist/skillhub, adding slug, displayName, version, and summary only to
the staged frontmatter. Canonical Skill content and its MIT-0 license are preserved.
The workflow retains staged files as an artifact; credentials are stored outside
that directory and removed after the run. See the
official CLI publishing guide.
A successful upload can return pending_review; public visibility requires
platform review. Existing-version or ownership errors fail the job; use a new
manifest version for changed content and verify the publishing account owns the slug.
GitHub Pages must be set to Build and deployment > Source: GitHub Actions.
skills.sh has no publish API — it indexes public repositories and ranks them
from anonymous skills CLI telemetry — so CI verifies public installability
and never generates install telemetry.
OpenAI and Claude public directories require human review; CI builds and
validates their submission artifacts but does not bypass those review portals.
See SUBMISSION.md.
The plugin ships no public MCP server. It connects to the user's local DCC-MCP
gateway through dcc-mcp-cli; a public multi-tenant MCP endpoint is a separate
security and deployment product.
Product catalog, package, and CI evidence proves discoverability and artifact parity only. It does not claim licensed real-host validation for any listed application.
Repository code is MIT. Published Skill content is MIT-0 as declared by each
SKILL.md and required by ClawHub.