add_to_board
ChatGPTAdd catalog items to one of the account's boards (id or slug). Idempotent: an item already in the board is reported under already_in_board, without adding it again. Call only when the user asks to add items. Returns the canonical URL, title, item count, and cover thumbnails.
analyze_palette
ChatGPTAnalyze any hex list (e.g. a brand's extracted colors) with the same detail the Color Palettes catalog carries: WCAG 2.1 contrast pairs, swatch roles (background/surface/primary/accent/text/...), temperature, harmony, hue families, saturation/lightness stats, color descriptors, and nearest named colors. On invalid input, returns {"error": "..."} instead of raising.
create_board
ChatGPTCreate a board, optionally seeded with catalog item ids (from search_inspiration/similar_items results). Soft-dedups by title: if the account already has a board with the same title (trimmed, case-insensitive), returns THAT board with a note instead of creating a duplicate. Returns the canonical URL, title, item count, and (when seeded) cover thumbnails.
download_image
ChatGPTReturn an item's image bytes as base64 JSON — ONLY for clients that cannot render MCP image content. WARNING: the base64 lands in model context (hundreds of KB at large sizes) — prefer get_image, which returns a native image block, and keep max_px small here.
fetch_image
ChatGPTDeprecated — call get_image instead (it also takes a list of ids). Kept only so existing integrations keep working.
find_starter_kit_page
ChatGPTFind the matching VASA Starter Kit page for a library subtype (section_type/slide_role/page_role/composition). Returns match = direct | closest | none.
generate_palette
ChatGPTGenerate a Relume-style 7-step tint/shade ramp from a brand hex color. color brand hex, e.g. "#2E8B57". mode "clone" (Relume-exact linear-RGB) or "perceptual" (OKLab, even lightness, stable hue). neutral_tint_strength 0..1 — how strongly to tint the neutral gray ramp toward color (0 = pure gray). find_similar how many nearest curated Color Palettes to return (0 to skip). Returns {color, mode, ramp[7], neutrals[7], contrast, roles, exports:{json,css,tailwind}, similar_palettes[]}. ramp/neutrals items are {name, group, hex, rgb}. contrast = WCAG 2.1 pairs over the ramp (best_text_pair, aa/aaa). roles = per-step role label. similar = nearest color-palettes items by ΔE. On invalid input, returns {"error": "..."} instead of raising.
get_board
ChatGPTOne board's metadata + items (id or slug accepted): title, description, public/private, curator, and compact item rows (id/title/category/subtype/image_url). Works for the caller's own boards and for anyone else's PUBLIC boards.
get_image
ChatGPTShow an item's image — or up to 6 of them in one call. Pass a single id for one image, or a LIST of ids to build a contact sheet without one call per image. Batched results come back as a manifest block first ({order, errors}) followed by the images IN THAT ORDER, so you can map each picture back to its id; an id that failed appears only in errors. Resolution defaults to 1600px for a single image and 1024px for a batch; set max_px to override.
get_inspiration
ChatGPTOne item's full record — or several, in a single call. Pass a single id for one record, or a LIST of ids (up to 20) to fetch a whole set at once instead of one call per item. Scalar in, scalar out; list in, you get {returned, items, missing} with items in the order you asked and missing naming any id that wasn't found. detail="standard" (default) is almost always enough: frontmatter facets + summary + body (capped) + public image/video URLs, with the huge payloads slimmed (website-styles design_system reduced to colors/typography/spacing/ theme). detail="full" returns everything and can be tens of KB per item. similar=N also attaches the N most similar items to each record, so you don't need a follow-up similar_items call.
kit_target_for
ChatGPTMap a library subtype (section_type/slide_role/page_role/composition) to the matching VASA Starter Kit page, with match = direct | closest | none.
list_facets
ChatGPTThe controlled vocabulary you can filter on: facet values + counts for subtype, style, color, mood, typography, industry, elements, layout and components. Use it to turn a vague brief into real filters. Scope it with category — library-wide counts are much larger and rarely what you want. Only the top most common values per field come back (default 25); truncated reports how many were cut, so treat the list as "most common", not "all". tags is open-vocabulary and tens of thousands of values wide, so it is left out by default — ask for it explicitly with fields=["tags"].
list_my_boards
ChatGPTList the connected account's boards: id, slug, title, item count, public/private, and a canonical URL for each (the public /b/<slug> share page when published, else the private account editor).
render_specimen
ChatGPTRender arbitrary text set in a font family (category=fonts) -> a PNG (base64 data + data_uri). Use to preview how a real brand name or headline looks before committing. Returns an error for Google Fonts stubs (no local binaries).
save_to_board
ChatGPTSave designs to one of your boards, creating it if it doesn't exist yet. Call only when the user explicitly asks to save designs. Pass ids from a search, or query/filters/category to search and save in one step. Reuses a matching title when found; lookup failures or concurrent creation can create duplicates. Existing items return under already_in_board. Returns the board URL, item count, and cover thumbnails.
search_inspiration
ChatGPTSearch design inspiration. Hybrid-ranked (text relevance + semantic), diversified, and compact by default (id, title, category/subtype, image_url, score, why) — pass verbose=true for the full facet arrays. To SHOW a result, call get_image(id): search returns URLs, not pixels. More without repeats: pass offset, or exclude_ids with what you showed. limit=1 answers "just give me one" and adds alternates. query natural language, e.g. "dark editorial SaaS hero with oversized type". filters exact-match AND across keys, any-of within a key. Keys: category, subtype, theme, asset_status, style, color, typography, mood, industry, tags, layout, components, elements. (subtype is section_type for website-sections [hero, cta, pricing, footer...], slide_role for slide-decks, asset_type elsewhere; for color-palettes, temperature/harmony/hue-families live in tags.) Call list_facets to see the real values before guessing. brand on-brand re-ranking: a website-styles id/slug (e.g. "elevenlabs") or a profile dict (style[], color[], mood[], typography[], industry[], tags[], palette[] hex, theme). diversify near-duplicate suppression (same deck/brand/look) — leave on unless you specifically want the clones. Zero results come back with suggestions — retry with one of those.
search_x_bookmarks
ChatGPTSearch the design references YOU bookmarked on X (Twitter) and imported into your Inspiration account (Boards page -> Connect X). Matches tweet text and author; empty query returns your most recent imports. Results carry image URLs, the live post link, and the saved tweet text. Private to your account — other users' bookmarks are never searchable.
similar_items
ChatGPTItems most like the given one: shared palette (CIE76 delta-E), facets, category/subtype/theme — diversified so results aren't one deck or brand. Compact rows by default (id/title/category/subtype/image_url + similar_score); verbose=true adds the full facet arrays.
suggest_pairing
ChatGPTSuggest complementary font families to pair with this one (e.g. a serif body for a sans display), ranked by mood/style/tags overlap across the opposite classification.
widget_item_detail
ChatGPTInternal backend for the moodboard widget's detail panel. Not for direct use — call get_inspiration / similar_items instead.