English | 中文
ModelScope Skills equip AI agents with the ability to operate on ModelScope — model/dataset management, Studio deployment, MCP service configuration, and Skills Center operations. It serves both ModelScope sites: the domestic site modelscope.cn and the international site modelscope.ai. Compatible with Cursor, Claude Code, Codex, Gemini CLI, and other major coding agent tools.
This repository follows the Agent Skills standard format.
Note
Start by installing ms-hub — the unified entry point for the entire ModelScope platform, enabling your agent to perform most platform operations via the ms CLI, Python SDK, and OpenAPI.
| Name | Description | Docs |
|---|---|---|
ms-hub |
Unified operations hub — model/dataset search/download/upload, repo management, Studio quick-ops, MCP services, Skills Center | SKILL.md |
ms-studio-deploy |
Deploy local projects to Studios (Gradio / Streamlit / Docker / static), API-first workflow with auto-diagnosis | SKILL.md |
MCP service management and Skills Center publishing are now folded into
ms-hub: common ops live in SKILL.md §8/§9, with depth inms-hub/references/mcp-services.mdandskills-center.md.
pip install modelscope
modelscope skills add @ModelScope/ms-hub
# If the modelscope/ms entrypoint has no `skills` command (shadowed by modelscope_hub), use the more reliable:
# curl -fsSL https://modelscope.cn/skills/install.sh | bash -s -- @ModelScope/ms-hubSkills install to ~/.agents/skills/ by default, where major agent tools auto-discover them.
Alternatively, install from source:
git clone https://github.com/modelscope/modelscope-skills.git
cp -r modelscope-skills/skills/ms-hub ~/.agents/skills/Copy the skill folders from skills/ into your project's .cursor/skills/ directory; Cursor auto-discovers the SKILL.md files there.
/plugin marketplace add modelscope/modelscope-skills
/plugin install ms-hub@modelscope/modelscope-skills
Copy or symlink skill folders into $HOME/.agents/skills (or $REPO_ROOT/.agents/skills).
gemini extensions install https://github.com/modelscope/modelscope-skills.git --consentOnce installed, just tell your agent what you need:
- "Download Qwen2.5-7B-Instruct"
- "Deploy this project to ModelScope Studios"
- "Find a map MCP service and configure it in Cursor"
- "Publish my Skill to the Skills Center"
The agent automatically loads the relevant Skill and executes.
pip install modelscope
export MODELSCOPE_API_KEY="your_token" # from the target site's /my/myaccesstoken
# Optional — pick a site (default: domestic modelscope.cn):
export MODELSCOPE_ENDPOINT="https://www.modelscope.ai" # international; omit for domesticModelScope runs two sites with separate accounts, tokens, and catalogs:
| Site | Endpoint (MODELSCOPE_ENDPOINT) |
Token page |
|---|---|---|
| Domestic (default) | https://modelscope.cn |
https://modelscope.cn/my/myaccesstoken |
| International | https://www.modelscope.ai |
https://www.modelscope.ai/my/myaccesstoken |
MODELSCOPE_ENDPOINT reroutes the OpenAPI base, CLI, SDK, and git URLs in one shot. Access tokens are site-scoped — use the token issued by the site you target. The skills detect the intended site from MODELSCOPE_ENDPOINT / your token / explicit intent and default to the domestic site when unspecified. See ms-hub → "Site selection & endpoint routing".
ms-hub (unified entry / quick reference)
├── Hub model/dataset/repo operations
├── MCP service search·deploy·configure → references/mcp-services.md
├── Skills Center search·install·publish → references/skills-center.md
└── ms-studio-deploy (full Studio deployment workflow, API-first)
ms-hub covers common operations as quick-reference commands; MCP and Skills Center depth lives in references, while the multi-step Studio deployment is split out into ms-studio-deploy.
- Create a new folder under
skills/ - Write
SKILL.md:--- name: my-skill description: What it does and when to activate --- # Title Guidance + examples + guardrails
- Add scripts, templates, reference docs as needed
- Submit a PR
Apache License 2.0