UXLoom

Your generator gave you 6 screens. UXLoom proves you're missing 9 states.

AI generators (v0, Lovable, Figma Make, Claude) produce happy-path screens. UXLoom is the critic layer: it models user journeys as state machines, treats screens as nodes with state contracts, and mechanically proves what's missing before a line of production code exists — unreachable screens, dead ends, missing error/empty/loading states, WCAG contrast failures, undersized touch targets, and labels that will overflow under localization.

Agent-native by design: the interface is an MCP server (works with Claude Code, Codex, and any MCP client), with Agent Skills included.

Deterministic by design: same input, byte-identical report — benchmarked (packages/bench) at 1.000 precision/recall on a seeded defect catalog, SHA-256-stable across processes, 1000 screens in under 5ms. That's what lets design completeness gate CI, where an LLM opinion can't.

Website: uxloom.dev · npm: uxloom · MCP registry: io.github.uxloom-dev/uxloom ·

uxloom check finding 9 errors in a generated checkout flow, then passing the repaired one

Packages

New here? Start with the Quickstart — prerequisites, the Claude Code walkthrough, what to say to your agent, and troubleshooting.

Connecting UXLoom to Claude Code and running the first check

Quick start (agents)

# Claude Code
claude mcp add uxloom -- npx -y uxloom

# Codex CLI
codex mcp add uxloom -- npx -y uxloom

The project file (uxloom.project.json) lives in your workspace and belongs in git — the design is data, versioned next to the code it specifies.

Quick start (humans & CI)

npx uxloom init      # one-command setup: MCP config + agent skill + starter file
npx uxloom preview   # live mocks (themed, commentable, EDITABLE in the browser)
npx uxloom export    # shareable HTML — plus --svg (Figma/Penpot import; add
                     #   --manifest for the round-trip key) and --png (playwright)
npx uxloom check     # design completeness — exit 1 on errors, CI-ready
npx uxloom audit     # implementation drift — web AND native (Swift/Kotlin/
                     #   Dart/Java markers); --live verifies the real DOM;
                     #   --design <file|dir> audits a Figma/Penpot export vs the contract
npx uxloom diff      # human-readable design diffs for PR review

Evidence-based design: every decision in the contract can carry its rationale — reasoning, rejected alternatives with pros/cons, sources, confidence — enforced by the critics once adopted, iterated through a bounded design_review loop (max 3 rounds), and shown to stakeholders in the preview's evidence panel (ⓘ) and exports. The design doesn't just validate; it argues its case.

Agent-addressable comments: a reviewer drops a pinned comment in the preview and clicks "→ agent". The comment becomes a work item any Gen-AI model can read with full context — comment_context returns the pinned layout block, the screen contract, the journey references, and the current findings for that screen — act on, and resolve back into the preview with a note. One click from feedback to addressed.

CI-native: check and audit take --json, --sarif (GitHub code scanning), and --github (inline PR annotations). Brownfield-ready: --update-baseline freezes existing findings so only new drift blocks; uxloom.config.json tunes thresholds to your accessibility bar. Full documentation: uxloom.dev/docs.html.

Add it to CI and a happy-path-only design can never merge:

- run: npx uxloom check design/uxloom.project.json

Workflow (also shipped as a skill in packages/mcp-server/skills/): project_init → brief_start/brief_answer → journey_define → screen_register → project_validate → fix → repeat until zero errors → coverage_report.

Does it actually catch things?

tools/dogfood.mjs drives the real MCP server through three products, twice each: screens as a happy-path generator hands them over, then repaired using the validation report. Artifacts in examples/.

Caught: an unreachable promo screen, dead-end verification states, five undesigned payment/error states, a 2.4:1 contrast button, a 40px touch target on Android, a checkout label that breaks in German, and three products' worth of missing offline states. Zero errors and zero warnings is reachable honestly — screens declare documented exemptions where a baseline state genuinely cannot apply, and contradictory exemptions are flagged.

Development

npm install
npm run typecheck
npm test

Status

Released and maintained: on npm and the official MCP registry, with the benchmark scorecard published in every GitHub release. The JourneyGraph format (formatVersion: "0.1") may evolve until 1.0; releases follow RELEASING.md — every surface is drift-checked in CI.

License

MIT