UXLoom is a UI/UX design validation tool for AI-generated interfaces. Its core idea is to act as a "review layer": AI generators (such as v0, Lovable, Figma Make, Claude) usually only produce happy-path interfaces, whereas UXLoom, before product code is written, models user journeys as state machines, treats interfaces as nodes with state contracts, and mechanically proves gaps and defects in the design.
Main capabilities
- Journey completeness validation: detects unreachable interfaces and dead-end states.
- State coverage checking: finds missing baseline states such as error, empty, and loading.
- Accessibility and adaptability checks: WCAG contrast failures, touch targets that are too small, and label overflow after localization.
- Evidence-based design: every decision in a contract can carry a rationale, rejected alternatives with their pros and cons, sources, and confidence, and is enforced by the reviewer; iterate through a bounded design_review loop (up to 3 rounds), presentable to stakeholders in the preview's evidence panel (ⓘ) and in exports.
- Agent-addressable comments: reviewers place pinned comments in the preview and click "→ agent"; the comment then becomes a work item readable by any Gen-AI model. comment_context returns the pinned layout block, interface contract, journey references, and current findings for that interface. Agents can process it and resolve it back into the preview with annotations.
- Determinism: identical inputs produce byte-for-byte identical reports; on a seeded defect catalog, benchmark precision/recall are both 1.000, SHA-256 is stable across processes, and processing time for 1000 interfaces is under 5ms. This makes design integrity usable for CI gating in a way that LLM opinions cannot.
How to use
- The project file uxloom.project.json lives in the workspace and is committed to git. Design as data, versioned alongside code.
- Workflow (also provided 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.
- For agents: the interface is an MCP server, compatible with Claude Code, Codex, and any MCP client.
- For humans and CI: check and audit support --json, --sarif (GitHub code scanning), and --github (inline PR comments). Brownfield support: --update-baseline freezes existing findings and only blocks new drift; uxloom.config.json can adjust thresholds to match accessibility standards.
Applicable scenarios
- Add a gate in CI so that designs containing only happy paths cannot be merged.
- Discover missing states, accessibility issues, and localization risks during the design review stage.
- One-click closed loop from feedback to resolution in agent-assisted design workflows.
Limitations and maintenance
- The JourneyGraph format (formatVersion: "0.1") may evolve before 1.0.
- Published and maintained on npm and the official MCP registry; the benchmark scorecard is announced with each GitHub release.
- Detailed installation commands, quick-start steps, and troubleshooting documentation are on the official website uxloom.dev and its documentation page uxloom.dev/docs.html.
