Search award-winning websites on Awwwards — the site index with filters. Returns site cards with inline screenshots, live URLs, awards and tags. Filters combine as AND. count defaults to 6 (max 12); page is one-based over the matched local set, and deep upstream query-string pagination is not crawled. Color searches always fetch the live filter page (site rows carry no colors). Use search_directory for agencies/studios, search_elements for UI components, get_collections for moodboards, new_winners for today's delta, and get_site_details for one site's DNA. A failed top-up keeps the already-cached page window; if the required live fetch fails, stale cached rows are served (page 1; color unchecked).

Get the design DNA of one Awwwards site: color palette, technologies, design elements, awards, description and inline screenshot. compact=true returns a token-diet summary. Batch: pass slugs (1–10, no images) for ordered compact details per slug with cached/fetched/error status; compact is ignored in batch mode (batch output is always compact). Provide exactly one of slug or slugs.

Compare the design DNA of 2–3 Awwwards sites: titles, live URLs, palettes, technologies, elements, awards and jury scores. Text only; missing details are fetched and cached. For your own build vs a reference, use verify_build instead.

Synthesize a cited design direction from 2–5 Awwwards site slugs plus a project goal. Uses cached detail (including its element list) and Motion DNA evidence first and fetches only missing details; includeStructure=true also runs a browser band analysis per reference. Returns one Markdown text block containing versioned JSON: shared signals vs one-site outliers, evidence gaps, an evidence-backed interaction plan, and reduced-motion/content/keyboard guardrails. The structuredContent field mirrors that JSON. Use search_sites to find slugs; use compare_sites for raw side-by-side DNA and analyze_page_structure for a band map — this tool synthesizes a brief; it does not score a build or run a motion census.

Read offline index status: cached site count, progress, last successful finish and error, lock state and freshness. No live requests.

Aggregate metrics over the LOCAL cache only (zero network). ALL-TIME: site count, gallery element count and technology/tag frequency; elements.corpus discloses the whole locally cached element set (total, gallery, site, categorized/uncategorized, builtWithKnown, fetched date range) while elements.total stays gallery-only. WINDOWED by windowDays (default 90, 1-3650): cached details, palette colour frequency and jury score stats from at most the 500 most recently fetched detail rows, plus Motion DNA library frequency. Each metric carries coverage labelled 'locally cached subset'; windowDays/detailsCached/detailParseCap appear only on windowed metrics (sites/elements/technologies are explicitly all-time). top caps each frequency list (default 10, 1-50). Aggregates only - no site slugs or titles. Personal, non-commercial use.

Get the design-element highlights of one Awwwards site: component-level visuals (3D models, video content, mobile layouts, microcopy) with poster images inline and video URLs.

Search Awwwards UI components — award-grade footers, menus, loaders, transitions, micro-interactions… — across cached gallery records plus source:"site" records derived from new SOTD winners' own Elements sections. It does not crawl the /elements/* gallery (robots.txt disallows it). Usage: pass query text; category and stack combine as AND when both are provided and are applied across the whole cached corpus before the result limit; each result carries its element slug, source and parent project identifiers.

Get one Awwwards element in full: title, author, built-with stack, media URLs, related elements and the parent award-winning project. Use search_elements first for ids.

List the filter taxonomy available on Awwwards: award collections, color hexes, tag/technology slugs usable with search_sites, and the annual winner categories.

List the official Awwwards Annual Awards categories (Site of the Year, Agency of the Year, …) with nominee counts. Annual Awards only: this is not the daily site index. Annual data comes from Awwwards' official public read-only API for personal, non-commercial inspiration use; nominee details are hydrated on request. Media URLs are reference-only — nothing is downloaded.

Annual Awards only: list nominees in one official Awwwards Annual Awards category, hydrated on demand at 1 request/second (default 10, max 20 per call). Site summaries carry siteSlug and inLocalIndex, so an agent can jump to get_site_details when the daily index has the site. Personal, non-commercial inspiration use only; media URLs are reference-only and are never downloaded.

Annual Awards only: get one official Awwwards Annual Awards nominee in full: site entries carry date, description, designBy, country, image/video URLs; agency entries carry location, avatar, award tag counts and featured projects. Pass id (a UUID or full /api/annual/nominees/... path) or category+slug; category+slug may hydrate up to 20 uncached nominees at 1 request/second (~20s worst case), so prefer id when known. Personal, non-commercial inspiration use only; media URLs are reference-only and nothing is downloaded.

Browse Awwwards' curated thematic moodboards, not a search index. Without a slug, lists collections (paginated); with a slug, returns that collection's first page of site cards. Curated slugs look like 'brutalism'; member collections need 'owner/slug'. One page per call, 7-day local cache. Personal, non-commercial inspiration use.

Search the Awwwards directory of agencies/studios, not sites. Filters are path-based, so one of country/category/budget applies per request; query is matched client-side over the fetched page. Returns public profile cards (name, slug, location, website, award counts) — contact details are stripped. One page per call, 7-day local cache. Personal, non-commercial inspiration use.

Audit any live page as evidence, not a grade: runtime frame/long-task/layout-thrash measurement via an installed Chrome/Edge over raw CDP (no playwright), plus static animation anti-patterns scanned over the page HTML and up to 8 same-origin CSS/JS assets (bounded: 8 assets, 2MB each, 1 request/second shared limiter, default 10s timeouts). Asset refs resolve against the final post-redirect page URL; asset redirects leaving that origin are rejected. Findings are severity-sorted before the 50-item cap, so worst is always present. nextActions ranks the first three fixes. compact=true returns the url, a reduced perf block (fps, droppedFrames, longTasks, layoutThrashing), pre-cap antiPatterns counts (count, worst, high, medium, info), coverage, nextActions and note; omit it for the full findings and full perf. Degrades honestly - perf carries {error} when no browser is available and the static scan reports its own error; neither suppresses the other. Personal, non-commercial use; audit only URLs you are authorised to inspect.

Collect accessibility EVIDENCE from a live page — an evidence census, not a compliance audit. Samples up to limit (default 100, max 200) visible text nodes with computed colour and a background composited bottom-up from the ancestor stack (colour alpha layers), then the contrast ratio (worst pairs with selector paths). Samples under any ancestor opacity group (<1) are excluded from ratios and counted in coverage.skippedOpacityGroups, never approximated; background images, gradients and pseudo-elements are not composited either, so contrast stays heuristic there. Also: image alt coverage; heading ladder (skipped levels, multiple h1); focusable/positive-tabindex counts plus outline:none stylesheet rules; unlabeled form controls; document lang/title. Result carries coverage and heuristic:true. http(s) URLs only; requires the optional playwright dependency and degrades to {error}. Personal, non-commercial use.

Take a fresh full-page screenshot of a live website URL using a headless browser. Requires the optional playwright dependency. Caveat: inner-scroll pages (body pinned, one container scrolls) return mode:"anchors" — six ordered viewport screenshots as JSON paths, not one full-page image.

Extract a page's section band map (tag, label, background color, offset, height per band) via a headless browser. Works on live URLs and file:// paths — use it to compare a reference site's structure against your local build. compare:"mobile" also captures at 390×844 and returns mobileBands + diff {heightDelta, bandsAdded, bandsRemoved, reordered} plus responsive evidence: horizontal overflow, small text and small tap targets. cdpEndpoint instead attaches to a browser you started (AWWWARDS_MCP_ATTACH=1) and scans the already-open matching page: it never navigates, scrolls, clicks or creates tabs.

Record a short motion-through video of a live website - preloader, scroll-triggered and hover/cursor animations - and return an inline filmstrip JPEG plus the .webm path. Requires the optional playwright and ffmpeg-static dependencies. Also writes a Motion Evidence Sheet (keyframes with census-stamped captions, diff masks, cursor-trail overlay, timeline strip) beside the filmstrip. This is not read-only: to expose hover/click motion it can click interactive elements and submit their forms.

Get the Motion DNA of a website: animation stack, ScrollTrigger/pin/scrub counts, easing vocabulary and duration distribution. Live captures append a censusSummary block (per-source scroll-census counts) from the same single launch. Serves the corpus first; live-captures unseen URLs (headless browser required).

Per-element motion census of a live page: what animates, its CSS properties, duration, delay, easing, and what triggered it (GSAP ScrollTrigger, native CSS scroll()/view() timelines, data attributes, SMIL/SVG). Includes svgStrokes list for stroke draw-in effects. Runs on any URL you give it (builds, references) — does not fetch awwwards.com. Run twice (with reducedMotion) to audit prefers-reduced-motion degradation. Optional scrubProbe drives the page and returns real progress curves per scrubbed trigger. compact=true returns a token-diet summary. Every payload also carries a choreography block: stagger hints (hint selector/value/source, groups) distilled from GSAP tween vars and stepped delays across sibling animations.

:hover interaction census of a live page: counts hover rules, samples up to 20 interactive elements, and reports which CSS properties change on hover with transition duration and easing. Use it to reproduce micro-interactions and verify interaction-state coverage. On mobile the hover loop is skipped and a note is returned. compact=true returns a token-diet summary. Runs on any URL — does not fetch awwwards.com.

Bundle forensics for a live page: fetch it once, download its script chunks (script tags, inline scripts, and modulepreload/preload-as-script hints; bounded: ≤40 chunks, 10MB, 1 request/second), save them locally, optionally de-obfuscate with webcrack if installed, and grep keywords across the code. Returns chunk paths + match lines so an agent can read the actual animation mechanism instead of guessing. Opt-in bundleIntel adds route, vendor, source-map presence and module-graph evidence; source maps are detected but never fetched, and default output keys are unchanged. Runs on URLs you are authorized to analyze — does not fetch awwwards.com.

Offline API-surface evidence from a local HAR 1.2 file (browser devtools export): groups requests by method + origin + path template (ISO dates, versions, UUIDs, hex ids and numbers templated; emails, tokens and long high-entropy segments redacted), with counts, statuses, content types and query names only. Reads the file only — never fetches recorded URLs, never returns bodies, headers or query values. No network.

Step-aligned comparison of interaction flows: runs the same 1–6 scroll/click/wait steps on a build and a reference page (two fresh headless launches, desktop or mobile), then compares path, scroll delta, skeleton hash, text hash and element count after each step. A step is same only when both sides are complete and untruncated; truncated or failed states are unknown. No screenshots: hashes are evidence, not visual truth. A click can still trigger client-side mutations. Runs on URLs you are authorized to inspect. Use verify_build for a scored structure/motion comparison; build_project converts a local file into a project and never verifies. Needs the optional Playwright dependency.

Convert a single-file exact rebuild (any local .html, e.g. the rebuild flow's full-page-exact.html) into a runnable Vite, Astro or Next.js project: sections become components, CSS splits per section, gsap timelines re-wire in the framework idiom (useEffect + gsap.context on Next), assets carried. Reads a local file and writes the project dir. The output is source, not a servable site: install and build it (npm install && npm run build) and serve the generated output (Vite/Astro dist, Next .next with next start), not the raw scaffold directory. Pairs with verify_build: build → convert → verify.

Compare a built site against a reference (an Awwwards winner, a client's old site, any URL): band-structure match + motion fingerprint similarity, with a pass/warn/fail verdict and evidence lines. Band labels are #id.firstClass, so give your rebuild's sections stable ids/classes; anonymous rebuild labels (empty/body, or duplicates) fall back to tag+height matching. compact=true returns the verdict, overall and band score, band counts (matched, refOnly, buildOnly), motion score and captureStable (false when either capture moved or the two captured different heights/animation counts); omit it for evidence lines. Live reports also carry a capture: evidence line with each side's rendered height, animation count and stability, because fixed waits can sample async content at different moments; a score that moved with capture state is visible there. Closes the design loop: build → verify → fix → re-verify. Reference can be an award site found via search_sites, compared read-only via compare_sites. scene3d is a SEPARATE opt-in 3D gate: it replays a capture_scene_3d pack on the build canvas, compares cropped PNGs pixel-by-pixel and returns its own status/verdict under scene3d (with diff PNGs), keeping overall/verdict legacy semantics; the legacy pass never certifies a canvas, and without pixelThreshold + maxChangedFraction it reports metrics only (inconclusive).

Serialize a page component (structure, computed-style tokens, animations) into a portable pack for rebuild tooling. Adds typography usage summaries (families/sizes/weights/line-heights/letter-spacings, top 12 each + sampled node count) and paletteSummary (colour counts from node styles + pack colors, transparent/rgba(0,0,0,0)/none filtered, top 24). compact=true returns a token-diet summary and omits both. Runs on any URL — does not fetch awwwards.com.

Extract a page's design tokens (colors, typography, spacing, radii, shadows, borders) from computed styles and return ordinal CSS custom properties plus a Tailwind theme map. Token names are ordinals (--color-1, --space-1, ...), never semantic guesses. Runs on any URL — does not fetch awwwards.com.

Image-delivery audit of a live page: totals by format, loading strategy, responsiveness and dimensions; flags broken, oversized, tiny and missing-alt images with issue samples. The page loads its own assets (no tool-initiated fetches) after load, a 3s settle, fonts and a pre-scroll that triggers lazy loading. Runs on any URL — does not fetch awwwards.com.

Capture a page's WebGL/canvas pose grid: samples the largest canvas (toDataURL, viewport-screenshot fallback) at up to 10 scroll steps with per-step scroll ratios, for studying scroll-driven canvas motion as frames. Frame evidence only — camera transforms are not recovered. Runs on any URL — does not fetch awwwards.com. Requires the optional playwright dependency.

Scene evidence in two strict operations. request.op:"capture" records a passive evidence pack for a WebGL/canvas page: at each checkpoint it scrolls the selected root, optionally moves the pointer, and saves a viewport screenshot plus a clipped canvas screenshot, with requested/actual scroll, canvas CSS rect/backing size/DPR and elapsed time. It also records a redacted inventory of the page's own responses (URL, MIME, size when declared, resource type, texture/model/HDR/shader/script/font/image guess, and an assetId per row) and passive framework hints (THREE.REVISION, BABYLON.Engine.Version, pc.version, matching script URLs, canvas count) without calling getContext or installing hooks. request.op:"download_assets" downloads up to 32 inventory assets from one capture: downloading is a licensing decision, so rights.basis (owned|licensed|permission) and rights.evidence are mandatory, are recorded with a timestamp in scene-captures//provenance.json, and are not verified by the server - do not download assets you may not use. It only fetches observed inventory URLs that carry no query string, fragment or credentials; there is no crawling, no guessed model or source-map URL and no credential use. Original bytes are written under scene-captures//assets with SHA-256 and an extension validated from MIME/magic (glb, gltf, bin, png, jpg, webp, ktx2, hdr, exr); unknown types are skipped unless allowUnknown saves them as .bin. Caps: 16 MiB per file, 64 MiB per run, 5 manual redirects, 1 request/second, one request at a time, 15s body deadline, 120s overall; every hop is DNS-validated and loopback/private/metadata/reserved addresses are refused. Honest scope: without the opt-in adapters no scene graph, camera, controls, shader source or geometry is recovered - frames show rendered output only. adapters are experimental and opt-in (at most two entries, each once): "three" reads objects reachable from window.THREE, window.scene/camera/renderer and window.app/experience/viewer/world after load (no heap crawl, no hooks, no network) and adds a bounded evidence.three summary (max 500 nodes, depth 20, 8 cameras, 8 lights, 32 materials, 200 texture URLs); "spector" needs the optional spectorjs package, installs its bundle as an init script before load and captures one WebGL frame into a bounded evidence.spector summary (command names/counts, context attributes, analyses, max 8 programs, 16 shaders, 64 KiB per shader source and 256 KiB of shader source total; command arguments, state, results, stack traces and memory are never transferred, and shader source is untrusted page text that is never executed). scene/camera/controls/geometry are marked available only for what three actually returned, shaders only when spector captured at least one source, and neither adapter promises scene recovery. Evidence is written to scene-captures//evidence.json plus PNG frames; responses are compact JSON with assetIds and downloadability, paths, hashes, counts, coverage and limitations, never inline image or asset bodies. The capture op requires the optional playwright dependency; download_assets does not.

Render a captured pose grid (pose-grid-v1 from capture_pose_3d) as a standalone scroll-scrubbed three.js stub page: frames preloaded inline, a synthesized camera path interpolated between snapshot scroll ratios (guidance only — the pose grid stores frames and scroll ratios, not recovered camera transforms), prefers-reduced-motion static fallback, optional census-derived st/pin/ease suggestions. Pure build — no page fetches, no awwwards.com access.

Fingerprint a page's framework (Next.js/Astro/Nuxt/Remix/SvelteKit/WordPress/Webflow/Framer), CSS framework (Tailwind/Bootstrap), and marker libs from fetched HTML — runs on any URL, one polite request, no browser

Generate a runnable GSAP+ScrollTrigger timeline stub from captured motion evidence (scroll census anims + optional synthesized camera anchors from a pose grid) — pure build from provided JSON, no page fetches

Build a motion.manifest.json scaffold from measured evidence: an analyze_page_structure result (bands) + an analyze_scroll_map census, with optional section overrides. Sections are mapped best-effort from each band's tag + first #id/.class token in its label (or sections[].selector) — no markup is available, so DOM initial-state hints are empty; census triggers still attribute to sections by those tokens. Returns { manifest, measured:false, censusAnims, warnings, note }; manifest.measured stays false when the census has no anims/triggers/curves. Never throws ({ error } on invalid input). Pure/offline — no page fetches.

Generate an offline Vite + Three.js scaffold from a capture_scene_3d evidence pack (scene-evidence-v1), not a recovery: the pack stores rendered frames and observed responses, so unless the experimental three adapter recorded a bounded scene summary, no scene graph, camera, controls, shader source or geometry is read, and captured application code, shader source, userData and callbacks are never executed. With evidence.three it recreates only Box/Plane/Sphere/Cylinder primitive meshes with position/quaternion/scale, MeshStandard/MeshBasic colour/opacity and locally bound texture maps, plus a static perspective camera applied from its exact captured matrix (matrix auto-update off); captured controls are never recreated, and unsupported geometry, materials, cameras and controls are listed as unresolved instead of guessed. Validates evidencePath itself (regular file, <=2 MiB, inside /scene-captures or a pack exported with top-level "exported": true) and up to 32 assetBindings (resourceId must be an observed inventory row that is not a script or shader; localPath must exist, be a regular non-symlink file <=64 MiB with a glb/gltf/bin/png/jpg/jpeg/webp/ktx2/hdr/exr extension; traversal and duplicate resourceIds are rejected). Returns files[{path,content}] (package.json with pinned three 0.160.0 + vite and dev/build/preview scripts, index.html, src/main.js with resize, reduced-motion static render, GLTFLoader for a bound model, texture loading, a clearly labelled placeholder plane when no model or captured mesh exists, disposal on pagehide and a load-failure message, scene.json describing bindings and frame references, and a provenance README), assetCopies[{from,to}] to public/assets/., unresolved[], authoredDefaults[] and provenance; mode is "reconstructed-subset" when a glb/gltf model is bound or the three adapter recorded at least one supported mesh, otherwise "placeholder". Deterministic and byte-stable, and strictly offline: no fetch, no install, no directory writes - copy assetCopies, then run npm install && npm run build yourself.

Search captured Motion DNA records: find sites that pin sections, scrub scroll-driven animation, or run a given animation library. Corpus only — live-capture happens via get_motion_dna.

Poll today's new Awwwards winners as a delta against the previous poll: new-entrant cards since the last call (first call seeds the baseline and reports no delta). Poll daily for a winners feed; pairs with watch_site for structured monitoring.

Longitudinal watchlist over Awwwards listings: track a studio (e.g. 'ToyFight'), a tag/technology slug (e.g. 'webgl'), or a site URL-slug for changes and new entries. add/list/remove; the delta is reported by watch_site list after new_winners or search_sites refreshes the listing.