# Capabilities & Tools **English** · [中文](../zh/capabilities.md) Qwen-MM-Plugins ships six capabilities, each installed separately (see [Installation](../../README.md#-installation)). This page lists what each provides. - [🖼 core — vision](#-core--vision) - [🎬 video-memory — long-video memory](#-video-memory--long-video-memory) - [✂️ video-edit — editing + generation](#-video-edit--editing--generation) - [🧊 blender — 3D modeling](#-blender--3d-modeling) - [📐 freecad — parametric CAD](#-freecad--parametric-cad) - [🎓 edu-agent — tutorial videos](#-edu-agent--tutorial-videos) --- ## 🖼 core — vision Foundational multimodal reading + analysis. Install name / entry: `qwen-mm-plugins-core`. **Reading (content fed directly to the model)** - `read_image` — dynamic-resolution image reading, aligned to the model's patch grid - `read_video` — extracts video frames with automatic FPS / resolution - `visualize` — general-purpose file visualization: PDF / Office / CSV / code / SVG / DrawIO / 3D / GIS / Notebook / LaTeX **Multimodal APIs (DashScope)** - `vision_chat` — vision chat, supporting image / video input - `ocr` — text recognition in images - `grounding` — object detection/localization, returning pixel bboxes (pairs with `draw_bbox` to draw the boxes) - `segmentation` — text-prompted segmentation (SAM3) - `transcribe_audio` — speech recognition (ASR), output as SRT / text / JSON **Image / frame output (saved to file + preview)** - `crop` — crop an image by box - `draw_bbox` — draw annotation boxes on an image - `save_view` — extract document pages / video frames into standalone image files **Web (Serper)** - `web_search` — web search, returning titles / snippets / URLs - `web_extractor` — fetch a web page's main text, optionally summarized - `image_search` — search by image (reverse image search) Keys: the multimodal APIs need `DASHSCOPE_API_KEY`; the web tools need `SERPER_API_KEY`. Native image/video/document reading needs no key. See [Dependencies](../../README.md#-dependencies). For exact schemas, see the capability's `SKILL.md` or each tool's inputSchema. --- ## 🎬 video-memory — long-video memory A hierarchical graph memory for QA over very long videos (30+ minutes). Install name / entry: `qwen-mm-plugins-video-memory`. Usage is workflow-based: ask `@/path/to/video.mp4 ` in your harness. On the first query the plugin builds a 4-level graph (Root → SuperEvent → MacroEvent → Subgraph) plus an embedding index next to the video, then answers using these query tools (a drill-down pattern): - `get_summary` — the video-level root summary (title, themes, key entities, tone) - `get_super_events` — list the high-level narrative arcs (super events) - `get_macro_events` — list macro events (optionally within a given super event) - `get_subgraph` — drill into one macro event's detailed subgraph (entities / events / edges / on-screen text) - `search_nodes` — semantic search over entity & event nodes by embedding similarity - `enumerate_events` — enumerate ALL matching event instances in time order (built for counting / "how many times" questions) - `search_ocr_text` — semantic search over on-screen text (OCR) nodes only - `search_asr_text` — semantic search over the speech transcript (ASR) nodes - `search_by_time` — find macro events covering a time range 📖 Full guide (install, env vars, examples): [video-memory-usage.md](video-memory-usage.md). --- ## ✂️ video-edit — editing + generation A video-editing skill plus DashScope generation tools. Install name / entry: `qwen-mm-plugins-video-edit`. **Generation tools (DashScope)** - `qwen_image` — image generation, editing, and translation (Qwen-Image) - `qwen_tts` — text-to-speech (Qwen3-TTS-Flash) - `wan_s2v` — digital-human lip-sync video (Wan2.2-S2V) - `wan_t2v` — text-to-video (Wan, wan2.7 series) - `happyhorse` — video generation and editing (HappyHorse) **Editing skill** — the editing side is driven by the skill under the capability's [`skill/`](../../src/capabilities/video-edit/skill/): `workflows/` (end-to-end recipes), `engines/` (rendering-engine selection matrix), `mcps/` (external-service catalog), plus `craft/`, `looks/`, and `review/` (technique, style, and QA references). Keys: the generation tools need `DASHSCOPE_API_KEY`. --- ## 🧊 blender — 3D modeling Drive a **running** Blender (3D modeling / materials / lighting / rendering) over a socket. Install name / entry: `qwen-mm-plugins-blender`. Thin client: the tools talk to a live Blender carrying the bundled blender-mcp addon. `QWEN_MM_AUTOLAUNCH=1` (preset in the plugin manifests) brings Blender up on the first tool call, auto-downloading it on Linux-x86_64 if missing; otherwise start it with `qwen-mm-plugins-blender --launch-app`. **Scene & code** - `execute_blender_code` — run arbitrary Python in Blender (the workhorse) - `get_scene_info` — summarize the current scene - `get_object_info` — inspect one object - `get_viewport_screenshot` — capture the viewport **PolyHaven assets** - `get_polyhaven_status`, `get_polyhaven_categories`, `search_polyhaven_assets`, `download_polyhaven_asset`, `set_texture` **Sketchfab models** - `get_sketchfab_status`, `search_sketchfab_models`, `get_sketchfab_model_preview`, `download_sketchfab_model` **Hyper3D / Rodin generation** - `get_hyper3d_status`, `generate_hyper3d_model_via_text`, `generate_hyper3d_model_via_images`, `poll_rodin_job_status`, `import_generated_asset` **Hunyuan3D generation** - `get_hunyuan3d_status`, `generate_hunyuan3d_model`, `poll_hunyuan_job_status`, `import_generated_asset_hunyuan` No API key needed to drive Blender (some asset/generation back-ends have their own keys, set inside Blender). 📖 Full guide (setup, `--launch-app`, env vars, troubleshooting): [blender-freecad-usage.md](blender-freecad-usage.md). --- ## 📐 freecad — parametric CAD Drive a **running** FreeCAD (parametric modeling, property edits, STEP/STL import/export, FEM). Install name / entry: `qwen-mm-plugins-freecad`. Thin client: the tools talk XML-RPC to a live FreeCAD carrying the bundled FreeCADMCP addon. `QWEN_MM_AUTOLAUNCH=1` (preset in the plugin manifests) brings FreeCAD up on the first tool call, auto-downloading it on Linux-x86_64 if missing; otherwise start it with `qwen-mm-plugins-freecad --launch-app`. **Documents** - `create_document`, `list_documents`, `reload_document` **Objects** - `create_object`, `edit_object`, `delete_object`, `get_object`, `get_objects` **Parts library** - `get_parts_list`, `insert_part_from_library` **Views & code** - `get_view` — screenshot a named standard view - `execute_code`, `execute_code_async` — run Python in FreeCAD **FEM** - `run_fem_analysis` — run a finite-element analysis (needs CalculiX) No API key needed to drive FreeCAD. 📖 Full guide (setup, `--launch-app`, env vars, troubleshooting): [blender-freecad-usage.md](blender-freecad-usage.md). --- ## 🎓 edu-agent — tutorial videos Turn a math/science problem (or an image of one) into a step-by-step Chinese explainer **video** or interactive page. Install name: `qwen-mm-plugins-edu-agent`. **Skill-only** — no MCP server. A pure Agent Skill: the model scaffolds, renders and voices the video itself (via `npx hyperframes` + Qwen-TTS). Because there is no MCP server, `uvx` installs nothing for it — its runtime deps (Node.js ≥18, hyperframes CLI, `dashscope`/`soundfile`/`numpy`/`requests`, ffmpeg, `DASHSCOPE_API_KEY`) are prepared manually. 📖 Full setup (dependency table, network boundary, prerequisites): [installation.md](installation.md).