Skill

Example Capability (template)

A minimal reference capability. To start a new one, copy capabilities/example/ and adapt it — see the repo's how to add a new capability doc for the full walkthrough.

You have the qwen-mm-plugins-example MCP tools available:

• echo — return text (a text-only tool).
• make_swatch — generate a solid-color image (an image return, built with shared.content.image).
• make_film_strip — generate N frames as images (the multi-frame / video pattern).
• describe — ask an OpenAI-compatible chat model, optionally about a local image (an API-calling tool; pass dry_run=true to see the request without a key or network).
• config_probe — report how config resolves via shared.env.get_env (env / config / default).

Check the tools in your tool list for full schemas. They exist to demonstrate the building blocks a real capability uses — content-return shapes, an API call, and env/config access.

Environment & config

Config is read via shared.env.get_env(name, default), with precedence environment variable > ~/.qwen-mm-plugins/config > default. For example, describe gets DASHSCOPE_API_KEY this way (through resolve_openai_endpoint), and config_probe reads the demo tunable QWEN_MM_EXAMPLE_GREETING. Run config_probe to see what currently resolves.

MCP-server code imports get_env from shared.env directly. Skill-side helper scripts run detached and cannot import shared — so if your skill ships its own scripts, copy the tiny env_config.py mirror (see video-memory / edu-agent) to get the same precedence.

Composing other capabilities (dependencies)

A capability should reuse other capabilities rather than reimplement them. This template has no media reader of its own, so:

• If the user gives a video, first call read_video from qwen-mm-plugins-core to extract frames, then feed a frame path to describe (to caption it) or make_swatch.
• If the user gives a document / PDF / 3D model, use qwen-mm-plugins-core's reading tools the same way before doing example-specific work.

When your capability depends on another, say so here and name the exact tool to call — that is how a skill tells the model to compose plugins instead of guessing.