ModelRunner Agent Skills

Portable AI-assistant skills for generating real product visuals — App Store screenshots and iOS demo assets — powered by ModelRunner, the unified AI inference API. Authored once as SKILL.md and built for Claude Code, OpenAI Codex, Cursor, and GitHub Copilot.

Each skill calls ModelRunner's public MCP server (mcp.modelrunner.run) — one API for image, video, audio, 3D, and text that you can use for anything else at modelrunner.ai »

Status: v1.0.0 — published and live on the Claude Code plugin marketplace as modelrunner/agent-skills (add it with the command below). Two skills across four tool targets, built from the two ModelRunner case studies.

What's inside

Skill	What it does
app-store-screenshots	Turns raw app captures into a cohesive set of App Store / Google Play marketing screenshots — device mockups, on-brand backgrounds, and ASO-aware headlines. Uses the modelrunner/app-store-screenshot-composer wrapper and the anchor→style-reference pattern for a consistent set. Case study »
ios-demo-assets	Bulk-generates on-spec demo images and loads them into Xcode's Assets.xcassets and the iOS Simulator's Photos library (xcrun simctl addmedia). Case study »

Prerequisites

Both skills call the ModelRunner MCP. Create a ModelRunner account with a small balance at modelrunner.ai, then connect the server at https://mcp.modelrunner.run/mcp and authorize with your account when prompted (OAuth). In Claude Code, installing a plugin below wires the MCP for you; for other tools the MCP config ships in dist/<tool>/.

Install

Claude Code

/plugin marketplace add modelrunner/agent-skills
/plugin install app-store-screenshots@modelrunner
/plugin install ios-demo-assets@modelrunner

OpenAI Codex CLI

SKILL.md is read natively. Copy the skills and add the MCP server:

cp -R dist/codex/skills/* ~/.codex/skills/
cat dist/codex/config.toml.snippet >> ~/.codex/config.toml

Cursor

Cursor reads the same Agent-Skills SKILL.md. Copy the generated .cursor/ into your project:

cp -R dist/cursor/.cursor /path/to/your-project/

GitHub Copilot (VS Code)

Copilot has no skill format, so each skill ships as a /-invokable prompt file:

cp -R dist/copilot/.github dist/copilot/.vscode /path/to/your-project/

Then run /app-store-screenshots or /ios-demo-assets in Copilot Chat.

How it's built

SKILL.md is the single source of truth (read natively by Claude Code, Codex, and Cursor — the "Agent Skills" open standard). A dependency-free build script emits each tool's packaging into dist/:

npm run build      # node scripts/build.mjs

• dist/claude/ — plugin marketplace (.claude-plugin/marketplace.json is at the repo root; plugin bodies + .mcp.json here)
• dist/codex/ — SKILL.md skills + a config.toml MCP snippet
• dist/cursor/ — .cursor/skills/ + .cursor/mcp.json
• dist/copilot/ — .github/prompts/*.prompt.md + .vscode/mcp.json

The name/URLs baked into the manifests come from package.json + mcp/modelrunner.json. Prefer a turnkey multi-tool sync instead of the built-in script? rulesync and ruler can emit the same targets from a canonical source.

License

MIT · Built by ModelRunner — the unified AI inference API.