Cookbook — Qwen-MM-Plugins Core

qwen-mm-plugins-core is the local file capability: it reads media at model-optimized resolution, visualizes documents and structured files, and writes crops, annotations, document pages, and video frames back to disk. It makes no cloud API call and needs no API key.

Cloud understanding lives in qwen-mm-plugins-api. Web and reverse-image verification lives in qwen-mm-plugins-search.

Tools

Read and inspect

• read_image — read an image at dynamic resolution
• read_video — sample video frames with automatic FPS and resolution
• media_info — inspect container, video, and audio metadata with ffprobe
• visualize — render PDF, Office, CSV, code, SVG, DrawIO, 3D, NIfTI, GIS, notebook, and LaTeX files

Write image views

• crop — crop a normalized 0–1000 box from an image
• draw_bbox — draw normalized 0–1000 boxes and labels on an image
• save_view — save document pages or selected video frames as standalone image files

For exact schemas, check the installed Skill or the MCP tool list.

Install

claude plugin marketplace add https://github.com/QwenLM/Qwen-MM-Plugins.git
claude plugin install qwen-mm-plugins-core@qwen-mm-plugins

core needs no API key. Video tools require ffmpeg/ffprobe; NIfTI visualization uses nibabel; other formats may need LibreOffice, TeX, Chromium, Blender, or FreeCAD. Run bash install.sh verify to check system dependencies for the installed capability.

NIfTI volumes

visualize opens .nii and .nii.gz files locally and read-only. For a 3D volume it returns metadata and axial, coronal, and sagittal center slices. For a 4D image it uses the first volume by default. Slices use the closest-canonical voxel axes; oblique volumes are not resampled. The source is never uploaded. This feature is for inspection and visualization, not clinical diagnosis.

@brain.nii.gz  Show its metadata and orthogonal center slices.

Cases

Case 1 — read a video, then extract Figure 2 from a PDF

This Claude Code session reads a full promotional video, opens a 35-page PDF, and saves a specific figure for closer inspection.

Shared Case: local views, cloud grounding, and web verification

The following Codex trace locates cakes, annotates the image, identifies a photographed place, and cross-checks the result on the web. It is shared with the API and Search cookbooks because the workflow crosses all three capabilities:

The trace was recorded before api and search were split out of core. Tool names carrying the old qwen_mm_plugins_core namespace map to the current capabilities shown above; the demonstrated workflow and outputs are retained because the session cannot currently be re-recorded.

Case 3 — ask a GUI harness to install Core and Edu Agent

The agent is asked to install core and edu-agent from this repository:

hello 帮我装一下 https://github.com/QwenLM/Qwen-MM-Plugins 的 core 和 edu 插件

This screenshot also predates the capability split. Current core advertises seven local MCP tools; cloud and search tools are installed separately.