media-gen-mcp

Generate or edit an AI image (text-to-image 文生图/AI画图; or image-to-image 图生图 via `images`) using free models (Agnes AI default, or Zhipu). Output downloads locally and the path is returned; no local rendering libs needed. WHEN: subject is photographic or illustrated (写实图/插画/概念图/original logo artwork…

Create an AI video (text-to-video / image-to-video / keyframe animation; 文生视频/图生视频/关键帧动画/让这张图动起来/做个动画) via free models (Agnes AI default, or Zhipu). Smart async: long videos return a handle to poll with `get_video`; short ones block until done. WHEN: user wants photorealistic or AI-generated video …

Poll and optionally download a video task created by create_video (by videoId or taskId). Companion to create_video — use it after an async video returns a handle, or to check/retrieve any video task.

Extract/recognize text from an image (OCR / 文字识别 / 文字提取 / 画像からの文字起こし) — verification codes, digits, license plates, printed Latin, or Chinese documents. Zero-config: tesseract runs in-process (WASM, bundled). Chinese accuracy is weak by default; configure a `paddleocr` provider for Chinese SOTA. The…

Recognize a table from an image → HTML/Markdown/JSON (表格识别/表格提取/tabla): invoices, receipts, financial statements, academic paper tables. Requires a `paddleocr` provider (PaddleX serving, Chinese SOTA); no pure-JS fallback (tesseract cannot parse table structure — a clear error is returned, not a sil…

Extract data points FROM an existing chart IMAGE (图表识别/图表数据提取/Chart OCR): reverse-engineer bar/line/pie/scatter charts into structured data series. Requires a `paddleocr` provider (PP-Chart2Table via useChartRecognition); if paddle is unconfigured, call `list_vision_capabilities` first to warn the u…

VLM image understanding — natural-language description or visual QA (图像描述/看图说话/VQA/describir imagen): handwritten text, complex layouts, scenes, formulas→LaTeX. Requires `paddleocr` provider (PaddleOCR-VL); M3+ adds `vlm` provider for enhanced VQA. Leave `question` empty for a default description. …

List available AI image/video models and video constraints per provider (Agnes / Zhipu). Use to discover model names (e.g. cogview-4, agnes-video-v2.0) and allowed video frame counts before calling generate_image / create_video.

Introspect vision (image recognition) provider capabilities BEFORE calling extract_text/extract_table/analyze_chart/describe_image — shows which providers are configured, what each supports (tasks/languages/latency/accuracy), per-task caveats, and routing guidance. Use this to avoid runtime errors (…

Generate architecture / flowchart / sequence / class / ER / mindmap diagrams (架构图/流程图/时序图/类图/ER图/思维导图/示意图), rendered locally to vector SVG. The D2 and Graphviz engines are BUILT IN (WASM, bundled with this tool) — you do NOT need d2/dot/graphviz installed, do NOT run `which d2`/`which dot`, and do N…

Generate a QR code (二维码) as SVG or PNG from text/URL. Pure local rendering — no qrencode/zbar/system install, no AI, no network. Just call with the text/URL to encode. NEXT: embed into a `generate_card` poster via the `logo` prop, or rasterize larger via `render_svg`. Multilingual triggers: QRコード …

Generate bar / line / pie / area / scatter charts and data visualizations (柱状图/折线图/饼图/散点图/数据可视化) from your data — Claude converts your numbers/CSV/data into a Vega-Lite spec internally; you just pass the data and chart type. Vega-Lite + vega are BUILT IN (bundled); no matplotlib / Python / graphviz …

Render a math formula to vector SVG (数学公式/公式渲染/方程). Pass the formula as LaTeX (e.g. E=mc^2, \frac{a}{b}, \sum_{i=1}^n i^2) — even simple formulas qualify; the user need not say 'LaTeX'. MathJax is BUILT IN (bundled) — no KaTeX/system install, no font dependency; just call this tool. Prefer this over…

Fetch and render a vector icon / logo / symbol / favicon (图标/logo/符号) from Iconify — 200k+ icons. Renders to SVG/PNG locally. Needs network (Iconify API); cached after first fetch. Browse at https://icon-sets.iconify.design. No AI. AVOID: original illustrated logo ARTWORK → use `generate_image` ins…

Generate a share card / OG image / quote card / poster / cover image (分享卡/分享图/封面图/海报/引言卡/金句卡/OG图; default 1200x630 PNG). The rendering engine is BUILT IN and runs entirely in-process — do NOT write HTML+CSS and screenshot it with headless Chrome/Puppeteer/Playwright, do NOT use Pillow/PIL/Python, an…

Render SVG source to high-quality PNG or SVG. Dual backend: resvg (92% filter fidelity, in-process) or Chrome (100% filter fidelity, needs system Chrome/Edge). AUTO-selects: SVG contains <filter>/<feGaussianBlur>/<feTurbulence> AND Chrome available → Chrome; else resvg. No AI; same input → same outp…

Render HTML/CSS/GSAP animation to a deterministic MP4/GIF/WebM video. Input: HTML source (with CSS animations or GSAP timeline) + fps + duration. Engine: headless Chrome (seek-based frame capture, HyperFrames-style) + ffmpeg (frame stitching). No AI, deterministic (same input → same output). WHEN: …

Extract text from a PDF document (PDF识别/多页OCR/财务报表/发票/扫描件文字提取): supports both digital PDFs (with embedded text layer → instant text extraction) and scanned PDFs (rendered to images → OCR via configured vision provider). Smart async: long PDFs return a handle to poll with get_pdf; short ones block un…

Poll and retrieve a PDF extraction task created by extract_pdf (by pdfId). Companion to extract_pdf — use it after an async PDF returns a handle, or to check progress of a long PDF extraction.