Palette Media

Image and video generation for Palette Code, as an MCP server and a CLI, on the Palette API (https://llm.pltt.dev).

• Palette Image: text-to-image and image editing in 8 steps at 2K, strong Korean typography, up to 10 reference images, transparent output. Sizes 1:1 4:3 3:4 3:2 2:3 16:9 9:16 or any WxH between 512 and 2752.
• Palette Video: 5–15 s clips with native audio. Korean on-screen text is rendered and OCR-verified (on_screen_text, up to 3 strings, clip regenerated until the text reads correctly). A reference image gives a reference-to-video clip.

npm i -g palette-media            # Node 18+. `npm i -g palette-code` pulls it in.
export PALETTE_API_KEY=...        # or sign in once with `palette-code auth login`

palette-media image "A4 poster for a Korean bakery opening, headline \"오픈 기념 20% 할인\", warm colours" --size 3:4 --out poster.png
palette-media edit poster.png "same poster, change the headline to \"주말 한정\"" --out poster-2.png
palette-media video "A Korean bakery storefront at golden hour, slow push-in" --seconds 5 --text "어서오세요" --out bakery.mp4
palette-media status <job-id>
palette-media watch https://www.youtube.com/watch?v=... --question "핵심 내용" --frames 12
palette-media transcribe meeting.m4a
palette-media info

• Understand a video: understand_video takes a file or a URL (YouTube and most sites via yt-dlp), returns title/channel/date/chapters, a timestamped Korean-first transcript (Palette STT), a visual timeline (frames read by the vision seat: scenes, on-screen text verbatim, charts) and a few frames to look at; transcribe_audio does the transcript alone. Needs ffmpeg and, for URLs, yt-dlp.

Palette Code runs palette-media mcp as the media MCP server. Tools: create_image, edit_image, create_video, video_status, understand_video, transcribe_audio, media_info. Files land in ./palette-media/ unless a path is given; every image answer carries a preview the model can look at, every video answer carries the verified-text report.

Generated media is AI-generated and the API response says so (ai_disclosure); keep the notice when publishing. MIT; see NOTICE.md.