The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Motionspec listing page.
MotionSpec is an open-core trust layer that checks and compiles reduced-motion-safe, on-budget UI animation for AI-generated web apps. An LLM authors a schema-validated JSON spec; a deterministic compiler emits vanilla-GSAP JavaScript + CSS — injection-proof and catalog-validated by construction, with an enforced prefers-reduced-motion fallback and a performance budget, with WCAG 2.2.2 (Pause, Stop, Hide) pause-path candidates reported — reduced-motion guards map to WCAG 2.3.3 (Animation from Interactions), Level AAA.
Run it two ways: as a keyless MCP server any LLM host can call (npx motionspec), or as a CLI compiler in your build (motion compile spec.json). Either way you keep plain files: vanilla-GSAP JavaScript plus CSS. (A WAAPI/CSS lowering exists in the codebase, but it has no CLI flag, no MCP tool and no schema target yet — see below.) MIT core. Docs: https://motionspec.dev
The thesis: capability lives in the catalog, not the model. A bigger model can write more elaborate specs, but it can never emit a primitive, parameter, or selector the Trust Boundary hasn't approved. The compiler trusts only what passes.
Not to be confused with: the Android Material Components
MotionSpecclass, the iOS material-motionMotionSpec, Motion.dev / Framer Motion, the usemotion.com calendar app, the Motion Specialties mobility brand, or text-to-video generators (Runway/Sora/Kling/Viggle). MotionSpec checks the UI animation inside web apps — it does not generate video.
The host LLM authors the spec; the Trust Boundary stays enforced either way. Listed on the MCP Registry as io.github.MasterPlayspots/motionspec. A hosted MCP endpoint is live: keyless motion_catalog + motion_validate at https://api.motionspec.dev/mcp (streamable-http; keyed tiers cover compile/audit/stats) — setup: https://motionspec.dev/docs.
This repo is also a Claude Code plugin: it bundles the MCP server (npx motionspec, all five tools, local, keyless) with two skills — /motionspec:motion (the author→validate→compile workflow) and /motionspec:audit <url> (motion-accessibility check, WCAG 2.2.2/2.3.3). Try it directly from a clone with claude --plugin-dir ., or install it from the community marketplace once listed:
| Version | v1.2.7 · schema frozen at spec v1 (ADR-0001, signed) |
| Published | npm motionspec (82 kB packed, 67 files, nothing dev-only ships) · MCP Registry |
| Tests | 302 green — injection attacks, 6000-spec fuzz, golden determinism, schema parity, pause-controls, motion-a11y audit · CI on Node 18/20/22 + x86 Playwright e2e |
| Catalog | 40 primitives, every one device-verified, reduced-motion-fallback mandatory; the 18 continuous loops also carry a WCAG-2.2.2 pause path |
| Supply chain | 2 runtime deps (MCP SDK, zod — both pinned) · 0 vulnerabilities · CycloneDX SBOM committed · all permissive licenses · CI actions SHA-pinned |
| Coverage | 99.06% lines / 99.09% functions / 79.05% branches of src/ (npm run coverage, 2026-09-04; CI gate 90/90/75) |
| Last audit | 2026-07-03 — 17/17 integration handshakes evidenced, infra 8.1/10, security: 0 critical, full-git-history secret scan clean |
| First client | CHS Computer — live on Vercel |
| Hosted MCP | live — keyless motion_catalog/motion_validate at api.motionspec.dev/mcp · keyed tier: Cloudflare Worker, per-key gated (hashed keys in KV) · two-stage rate limiting (pre-auth per IP + per key, burst-verified) · per-minute cron canary + external heartbeat (synthetic fault → alert in <5 min, proven on real infra) · Analytics Engine telemetry, PII-scrubbed · gated /dashboard |
Schema v1 is frozen: specVersion "1.0" is the stable public contract; "0.1" is deprecated and accepted until v1.2 (a tripwire test enforces the revisit). The [MS-XXX] error-code registry is public API — codes are never reused or redefined.
JSON.stringify) or a CSS-screened raw value through one shared safety gate (safety.js). Malicious model output is rejected fail-closed — tested and fuzzed over 6000 random specs.respectReducedMotion is default-on at the compiler level (fail-safe): omitting it still yields a prefers-reduced-motion guard. Opting out is possible but emits MS-GLOBALS-RRM-OFF; a prompt-side instruction alone can never disable the guard.a11y.persistent, and the compiler emits a pause path by construction: an animation-play-state: paused rule keyed on html[data-ms-paused] (outside the reduced-motion guard, so it is always live) plus, under pauseControls: "auto" (the fail-safe default), one accessible pause/stop toggle (type="button", aria-pressed in sync, ≥24 px target, visible focus ring, not rendered under reduced motion). pauseControls: "api" keeps the CSS contract and leaves the control to the integrator; "off" opts out but emits MS-GLOBALS-PAUSE-OFF when a persistent motion is present. The promote-gate refuses any infinite/repeat:-1 primitive that is not a11y.persistent. A spec with no loops adds zero extra bytes.catalogVersion for reproducibility (MS-CATALOG-PIN-MISMATCH fail-closed).model | model-repaired | cache-hit | escalate-* (local: JSONL sink · hosted: Cloudflare Analytics Engine, PII-scrubbed). Escalation clusters are the growth signal for new primitives.What you can get out of the compiler today, through the CLI or the MCP tools, is exactly one target:
vanilla-gsap — GSAP + ScrollTrigger. meta.target accepts nothing else (schema frozen at v1).A second lowering exists in the codebase and is kept green by the test suite, but it is not reachable through any interface:
src/compiler/lower-waapi.js) — zero-GSAP output on Element.animate, IntersectionObserver, and @keyframes/position: sticky. Full catalog coverage, byte-identical golden per primitive, same accessibility guard, same CSS safety gate. This is the framework-decoupling hedge: the IR outlives any animation library. Internal — referenced only by the tests and bin/promote-gate.js; there is no CLI flag, no MCP tool and no schema target for it (ADR-0001 freezes meta.target to vanilla-gsap; engine wiring is out of scope, ADR-0002). Do not plan a build on it until a release note says otherwise.The Catalog Forge (CI workflow, manual dispatch) picks the top telemetry-ranked gap, generates one candidate primitive, drives it through a multi-stage gauntlet — meta-schema, mandatory reduced-motion fallback, performance budget, output determinism (entropy tokens like Math.random/Date.now fail the gate), catalog-SemVer legality, golden creation — and opens a PR. It cannot merge, publish, or deploy: structurally (workflow permissions carry no packages/id-token, PR-only) and by regression test (forge-workflow-guard fails CI if anyone smuggles a publish step in). Gate 1 is always a human taste review.
| Tool | Contract |
|---|---|
motion_catalog | primitives + authoring rules + catalog version (16-hex pin) |
motion_validate | fail-closed Trust Boundary; precise [MS-XXX] errors; surfaces deprecations |
motion_compile | deterministic spec → code; never emits on a failed validation |
motion_audit | static motion-a11y check of a live URL (read-only, open-world) |
motion_stats | telemetry summary (escalations = catalog growth signal) |
Input is size-capped (MS-INPUT-TOO-LARGE, 64 KB). The stdio server exposes one tool factory as the single source of truth, contract-tested in test/mcp.test.mjs. A hosted MCP endpoint is live: keyless motion_catalog + motion_validate at https://api.motionspec.dev/mcp; keyed tiers cover compile/audit/stats.
motion audit <url> (CLI, --json for the machine payload) and the motion_audit MCP tool run a static scan of a page's HTML and linked stylesheets — no headless browser, no new dependency. It reports four motion problems: CSS animation/transition without a prefers-reduced-motion guard (WCAG 2.3.3), animated properties other than transform/opacity, infinite animations with no pause path (animation-play-state/data-*), and <marquee>/autoplay motion over 5 s (WCAG 2.2.2). Each finding carries a selector, the WCAG reference, and a copy-paste fix. It is honest about its limits: runtime motion (WAAPI/GSAP/JS) is reported as not audited (V2) rather than silently passed. A page that clears every check earns the reduced-motion-safe badge — the exact output MotionSpec itself produces.
motion audit --json is stable enough to gate a pull request. examples/ci/motion-audit.yml is a copy-and-adapt GitHub Actions workflow that builds your site, serves the build directory on localhost, audits the paths you list with npx -y -p motionspec@1.2.7 motion audit <url> --json (local, MIT, no key, no hosted call), and compares each page with a checked-in baseline .motionspec/baseline.json.
The gate fails when a page got worse — the same rule MotionSpec's weekly re-scan uses: the score fell, or the number of Level-A findings (WCAG 2.2.2 Pause, Stop, Hide) rose. A page without a baseline entry never fails; that run is the baseline. Re-baselining is a deliberate manual run (workflow_dispatch with update_baseline: true) that uploads the new file as an artifact for you to commit — the workflow never commits on its own. Fixed findings are listed as - fixed: lines, new ones as + new finding:.
The machine payload is { ok, url, score, badge, findings: [{ selector, rule, wcag, fix }], summary, disclosures }; badge is the literal "reduced-motion-safe" only at zero findings. Note that audit takes a URL, not a directory — hence the local server step. And the scope caveat travels with it: this is a static CSS scan (no inline style="", @import, CSS-in-JS, external JS bundles, video/GIF/Canvas, or flashing checks); a green gate means "no regression in the loaded CSS", not "accessible".
MotionSpec is a governed format, not just a tool. The normative spec is SPEC.md (versioned 1.0, RFC-2119 MUST/SHOULD/MAY over the JSON Schema, with a documented ADR-based change process). CONFORMANCE.md defines the five checks (schema, diagnostics, output, determinism, accessibility) an implementation passes to call itself MotionSpec 1.0 compatible, run against the published test/golden corpus. Multiple implementations passing the same corpus is what makes it a standard.
MotionSpec turns specific legal and normative accessibility requirements into compiler-enforced defaults. Each check maps to the frameworks that mandate it:
| MotionSpec mechanism | WCAG 2.2 | EN 301 549 | U.S. Section 508 | EAA / BFSG |
|---|---|---|---|---|
Pause/stop path for every continuous loop (animation-play-state + accessible toggle) | SC 2.2.2 Pause, Stop, Hide (Level A) | clause 9.2.2.2 (mirrors the WCAG SC) | incorporated (WCAG 2.0 A/AA baseline; 2.2.2 is Level A, in scope) | conformance presumed via EN 301 549 |
Reduced-motion guard on every motion (prefers-reduced-motion) | SC 2.3.3 Animation from Interactions (Level AAA) | clause 9.2.3.3 | beyond the AA baseline; provided anyway | supports the EAA "perceivable/operable" duties |
motion audit static checks (loops without a pause path, motion missing a reduced-motion guard, autoplay > 5 s) | 2.2.2 / 2.3.3 | clause 9 (web) | WCAG-incorporated success criteria | pre-market self-check for covered products |
Notes: EN 301 549 is the EU harmonised standard whose clause 9 adopts the WCAG success criteria by number. U.S. Section 508 (Revised) incorporates WCAG 2.0 Level A and AA — SC 2.2.2 is Level A and therefore in scope; SC 2.3.3 is Level AAA and is provided as a stronger guarantee than the baseline requires. The European Accessibility Act (EAA) and its German transposition (BFSG, applicable from 28 June 2025) require covered digital products and services to be accessible, with conformance commonly demonstrated against EN 301 549. MotionSpec enforces the motion subset of these obligations by construction; it does not by itself make an entire product conformant.
Live model instead of --mock: set MOTION_API_KEY (or OPENROUTER_API_KEY); optional MOTION_MODEL (default anthropic/claude-haiku-4.5) and MOTION_BASE_URL (any OpenAI-compatible endpoint). See .env.example.
Releases run the whole chain plus a canonical-clone guard and finish with a registry truth check — a version is "live" when the npm dist-tag says so, not when a local run went green.
Defense in depth on the hosted path: constant-time admin-secret comparison (no timing side channel on position or length) · customer keys stored hashed (SHA-256) in KV, fail-closed on any lookup error · pre-auth per-IP rate limiting closes the key-enumeration gap before auth work starts, per-key limiting after · throttled abuse alerts with zero PII · telemetry scrubbed before storage · strict CSP/X-Frame-Options/nosniff on the only ungated page (a data-free dashboard shell). Full posture incl. reporting: SECURITY.md. Last audit (2026-07-03): no critical findings, no secret ever committed across 197 commits of history.
.cursor/rules/motionspec.mdc and .github/copilot-instructions.md.docs/adr/0001-schema-freeze-v1.md — the frozen v1 contract and why.CONTRIBUTING.md covers setup, the gate-driven PR checklist, commit conventions, golden-file regeneration, and a short architecture tour. Issue templates live under .github/ISSUE_TEMPLATE/.
MIT.