awwwards-mcp v1.8.0

Deterministic design briefs and evidence-backed interaction guidance.

• Added the create_design_brief MCP tool for cited, versioned briefs from
  shortlisted Awwwards references.
• Reused cached site details, element metadata, and Motion DNA; optional
  structure analysis runs only when requested.
• Added an Interaction Plan for observed cursor/hover, scroll, entrance/reveal,
  easing, and duration signals, with reduced-motion and keyboard-access
  guidance. Missing evidence is reported rather than inferred.
• Documented the design-brief workflow in the awwwards-inspiration skill and
  README.

awwwards-mcp v1.7.2

Reliability & housekeeping. No runtime tool changes.

• Parser-drift probe: network rejections no longer masquerade as parser
  drift. A raw fetch() crash (DNS/reset/classifier network errors)
  killed the live probe with exit 1, which the workflow reported as
  "markup changed — parsers need updating" even though zero anchors had
  been probed. Rejected fetches are now recorded as failed pages and the
  verdict logic emits its existing honest fetch-fail (exit 2) state.
  This was the root cause of issue #6 (false alarm — the 9/19 committed
  status was ok 0/59, and awwwards.com fetches fine).
• Dependency alerts cleared: vitest + @vitest/mocker bumped to 4.1.11
  (dev-only; full suite 251/251 green on vitest 4). Supersedes the two
  Dependabot PRs, which were merged into one local commit.
• CI updated to actions/checkout@v5 / actions/setup-node@v5 in all
  three workflows (fixes the Node 20 deprecation annotation).
• Removed the stale elements-motion-dna worktree (fully merged).

awwwards-mcp v1.7.1

Daily environment self-check + corrected optional-dependency install hints.

• Daily env check: at server startup (once per day, stderr-only, never
  blocks serving) the server probes playwright with a real chromium launch
  and resolves ffmpeg-static from its own install location. When captures
  or motion recordings later fail, the cause was warned about up to a day
  earlier — and the fix hint prints once, not on every failed capture.
• Corrected install hints: capture/motion hints now say
  npm install -g playwright && npx playwright install chromium /
  npm install -g ffmpeg-static — the server resolves optional deps from
  its own install location, so a -D install in whichever project the
  agent happens to sit in does not fix a globally-installed server.
• Failure-tolerant by design: a flaky probe (first-run policy, AV scan)
  records nothing, so the next start re-checks; every failure is swallowed
  and stdout is never touched.

awwwards-mcp v1.7.0

First stable-minor release. npm versions 1.1.0–1.6.0 were the retired
alpha line, so the stable line jumps to 1.7.0 — semver forbids reusing a
published version number.

Winner feed (new_winners + watch_site), awwwards-setup skill, repo
renamed to the correct spelling.

• new_winners polls today's freshly-crowned winners (SOTD /
  Developer Award / Honorable Mention) against a persisted UTC-day
  baseline: the first call seeds silently and dumps the listing; later
  calls report only the delta. Each first-seen winner's site page is
  fetched once and its Elements section is backfilled into the searchable
  element corpus as source:"site" records (slugs namespaced
  site-<siteslug>-<title> so they never collide with gallery records) —
  new winners are element-searchable immediately, no gallery indexing
  needed. Per-winner fetch failures never abort the loop.
• watch_site adds persistent watches (studio / tag / url, optional
  award + note). list matches each watch against the freshest cached
  listing and reports per-watch NEW since last check deltas; it never
  fetches — new_winners / search_sites keep the pool fresh. First
  list seeds the baseline silently; url watches take the site slug;
  studio matching is a case-insensitive substring.
• New skill awwwards-setup — first-invocation onboarding: wires the
  MCP server when it isn't connected yet (Claude Code, Codex, OpenCode,
  mcpServers-standard clients, pi), asks the user's preferences once
  (result density full/compact, default viewport, live-capture opt-in
  with playwright install, local-index opt-in, winner watches to register)
  and persists them to ~/.awwwards-mcp/preferences.json. Later runs
  apply silently and only ask about missing fields. Shipped in the npm
  package; awwwards-inspiration hands off to it on first run.
• Docs: README documents both new tools and the two-source element
  corpus; awwwards-inspiration gains the tool-table rows, a "staying
  current" section, and the first-run setup handoff.
• Repo renamed Awwards-mcp → Awwwards-mcp (three w's, matching the
  npm package and awwwards.com). Old GitHub links redirect automatically;
  package.json and issue-template URLs updated.

awwwards-mcp v1.6.0

Skill-memory flywheel (reviewed and approved for ship):

• scripts/skill-memory.mjs (record/recall/distill/stats): the skills
  self-improve from what their verification loops catch. Findings journal
  (episodic, per-machine) + deterministic distiller folding rules seen 2+
  times into installed skill copies (machine-managed section, LRU-capped).
• promote + sync commands: promote prints folded rules cleaned for the
  shipped-copy PR (machine-local evidence paths stripped); sync copies repo
  skill updates into installed copies WITHOUT wiping learned rules.
• skills/_memory/techniques.json: techniques registry seeded from
  researched video-understanding, motion-detection, UI-structure,
  micro-interaction and image-analysis methods; grows via the flywheel.

awwwards-mcp v1.0.0-beta.1

First beta. Supersedes the alpha 1.x line — the full v1.6.0 feature set
(6 tools, local index, three skills incl. the self-improving memory flywheel,
self-heal doctor, drift monitor, dual-viewport capture) now carries a beta
label while it takes public hardening. Alphas remain on the registry for
reproducibility.

awwwards-mcp v1.4.0

Search intelligence (FTS5 — no new dependencies):

• Porter-stem + prefix matching with BM25 ranking: "magazines" now finds
  Magazine-tagged sites (68 on the live index), best matches first — the old
  substring path returned zero. "edito" finds "editorial"; multi-word queries
  keep AND semantics (every token must hit the same site).
• Zero-result queries gain loose OR-matched hints alongside the taxonomy
  suggestions. Live top-up merge and client-side filters unchanged.

awwwards-mcp v1.3.0

Mobile viewports:

• capture_live_site, analyze_page_structure and record_site_motion
  accept viewport: "desktop" | "mobile" (mobile = 390×844 @3x with
  isMobile + hasTouch; desktop 1440×900 default, byte-identical).
• Motion recordings render at the selected viewport — no more pillarboxed
  mobile filmstrips.

awwwards-mcp v1.2.0

Adoption kit:

• Demo GIF at the top of the README — real tool output (search results,
  design DNA, a motion filmstrip of our own build, a band map), generated by
  scripts/make-demo.mjs.
• Release automation: pushing a v* tag now gates on the CHANGELOG
  section + npm-absence, runs typecheck/tests/build, publishes to npm with
  NPM_TOKEN, and creates the GitHub Release from the CHANGELOG section.
• Package metadata: keywords, repository, bugs, homepage for registry
  discovery.