The AI Design Stack Just Split: Claude Design, Codex Sites, Figma, Lovart—and the MCPlato Control Room
A 2026 deep review of Claude Design, Sites in Codex, Figma, Lovart, and MCPlato—and why AI design now belongs in a web-experience stack, not one winner-takes-all category.
Published on 2026-06-05
The wrong question in 2026 is “which AI design tool is best?”
The sharper question is: which job are you hiring AI to do? A founder needs a landing-page concept; a product team needs a responsive web experience; an engineer needs a hosted demo; growth needs campaign visuals. “AI design” has split into visual exploration, hosted web creation, design-system governance, marketing creative, and control-room coordination.
That is why Claude Design, Sites in Codex, Figma, Lovart, and MCPlato belong in a stack, not one ladder.
Premium editorial hero showing an abstract responsive web experience emerging from a studio workspace
Figure 1: AI design is becoming a web-experience stack, not one universal canvas.
Naming reality check: Claude Design is real; Codex Sites is shorthand
Anthropic launched Claude Design on April 17, 2026 as an Anthropic Labs research preview.1 It can create designs, interactive prototypes, visual one-pagers, marketing collateral, and early landing-page concepts through conversation and direct refinement.1 That makes it a real 2026 product direction—but still a preview, not a replacement for professional design governance.
Codex Sites is useful shorthand, but the official naming is Sites or the Sites plugin inside Codex. OpenAI’s developer page is titled “Sites – Codex” and defines the feature as a way to “build and deploy hosted sites from Codex with the Sites plugin.”2 Sites lets Codex create, save, deploy, and inspect websites, web apps, and games hosted by OpenAI.2
The boundary matters. Sites is in preview, available for ChatGPT Business and Enterprise workspaces, with Business enabled by default and Enterprise controlled through administrator RBAC.2 Its output must be Cloudflare Worker-compatible ES modules, and every Sites deployment URL is a production deployment.2 Treat it as a hosted web-experience workflow, not final production design.
Product-by-product review
Claude Design is the strongest starting point when the experience is still vague. It can help a team explore what a homepage, onboarding flow, product narrative, or prototype could feel like before anyone commits to a design system. Its advantage is speed and conversational refinement. Its limitation is authority: it can suggest taste, but a team still needs accessibility checks, brand judgment, and downstream ownership.
Sites in Codex is most compelling when the desired artifact is an actual hosted web surface: a landing page, internal tool, dashboard, interactive demo, small web app, or game. It fits teams that already have structured requirements, code context, or an existing project they want to expose as a reviewable web experience. The OpenAI web-development use-case collection reinforces Codex’s role in web work, while the Codex changelog is the better place to track platform changes over time.34 Its weakness is that a production URL can create a false sense of completion. Deployment is not brand, legal, performance, or security approval.
Figma remains the professional source of truth for product and web design. Figma Sites gives design teams a web-facing creation path, while Figma Make adds prompt-to-app workflows with chat, attachments, point-and-edit, style context, a code editor, and publishing beta.56 Figma Design AI covers asset search, content replacement, interactions, layer renaming, text operations, image tools, vectorization, and First Draft; Figma’s AI agent began limited beta rollout on May 20, 2026.78 The web stack still depends on design-system discipline.9
Lovart is better framed as a creative and marketing visual agent. Its official page emphasizes natural-language pipelines, a unified canvas, image and video generation, layer-level editing, batch generation, and PNG/SVG exports.10 That makes it useful for campaign worlds, moodboards, hero-image directions, product-photo concepts, and launch variants. It is not the best place to govern a responsive design system or inspect a hosted site.
MCPlato fits the control-room layer. It is not a professional design canvas, a hosted-site platform, or a specialist marketing image studio. It is useful when the site project depends on research, source materials, long-running tasks, approvals, sessions, artifact discipline, and handoff. MCPlato’s public positioning around observable office AI work makes it a workspace for coordinating the job around Claude Design, Codex, Figma, and Lovart—not a replacement for them.11
Comparison table: what each tool should own
| Product / workflow | Primary job-to-be-done | Best at | Weak at | Best inputs | Output / artifact | Best user | Not ideal for | Recommended role in stack |
|---|---|---|---|---|---|---|---|---|
| Claude Design | Conversational visual exploration | Concepts, prototypes, landing-page directions, one-pagers | Final design-system authority | Briefs, product ideas, reference styles, direct feedback | Visual concepts, prototypes, draft experience layouts | Founders, PMs, designers exploring direction | Pixel-perfect production design | Experience ideation layer |
| Sites in Codex | Hosted web experience creation | Websites, web apps, games, dashboards, internal tools, demos | Visual taste, governance, review discipline | Requirements, code/project context, constraints, review criteria | Hosted website, web app, internal tool, dashboard, landing page, interactive demo | Engineers, technical operators, product teams | Unreviewed launches or brand-new identity work | Hosted prototype and web execution layer |
| Figma | Professional design source of truth | Design systems, collaboration, responsive web design, governed assets | Replacing human design judgment | Components, styles, product requirements, team libraries | Design files, prototypes, sites, branded assets | Design and product teams | Unstructured research orchestration | System-of-record design layer |
| Lovart | Marketing creative exploration | Moodboards, campaign visuals, brand directions, batch creative | Design-system governance, site inspection | Brand prompts, campaign goals, visual references | Images, videos, layered creative assets, exports | Marketers, creators, brand teams | Product UI source of truth | Creative exploration layer |
| MCPlato | Observable cross-material work coordination | Research, approvals, sessions, artifacts, async tasks, handoff | Specialist design canvas or hosted web runtime | Sources, local materials, task plans, approvals, briefs | Reports, outlines, task trails, deliverables | Operators, researchers, PMs, teams managing work | Replacing Figma, Codex, Lovart, or designers | Control-room layer |
Premium semi-3D architectural cross-section of an AI-assisted web experience workflow
Figure 2: The stack separates exploration, governed design, hosted execution, creative assets, and coordination.
Scenario recommendation matrix
| Scenario | Primary recommendation | Secondary tools | Why |
|---|---|---|---|
| Landing page visual direction | Claude Design | Lovart, Figma | Claude Design accelerates early exploration; Lovart widens mood and campaign directions. |
| Responsive site design | Figma | Claude Design, MCPlato | Figma should own components, responsive behavior, collaboration, and reviewable decisions. |
| Prompt-to-web app prototype | Sites in Codex | Figma Make, MCPlato | Sites in Codex fits hosted web apps or demos from technical context. |
| Internal dashboard/tool | Sites in Codex | Figma, MCPlato | Codex can produce a hosted technical surface; Figma governs UX patterns and MCPlato coordinates requirements. |
| Marketing campaign assets | Lovart | Figma Buzz, Claude Design | Lovart is better for creative variety, moodboards, and launch visuals; Figma Buzz helps bulk assets. |
| Cross-material research before site build | MCPlato | Claude Design, Figma | MCPlato is the center for source gathering, notes, approvals, and handoff. |
| Existing project to hosted demo | Sites in Codex | MCPlato | Codex Sites can turn compatible web work into a hosted deployment; MCPlato tracks caveats. |
| Design-system governed web experience | Figma | Sites in Codex | Figma owns the system of record; Codex can expose a compatible implementation for review. |
| Long-running site project coordination | MCPlato | Figma, Sites in Codex, Lovart | Multi-stage site work needs visible research, decisions, assets, deployment notes, and approvals. |
What teams should avoid
First, do not treat a hosted URL as final approval. OpenAI’s Sites documentation is explicit that every deployment URL is a production deployment.2 A working demo still needs source review, accessibility review, brand review, security review, and stakeholder sign-off.
Second, do not let shorthand confuse ownership. “Codex Sites” is a convenient title, but the official feature is Sites inside Codex. Figma should own governed design systems. Lovart should not become the source of truth for product UI. Claude Design should not be treated as finished production design. MCPlato should coordinate materials and decisions, not pretend to be a design canvas or hosting layer.
High-end editorial product photography comparing design canvas, hosted web object, creative studio, and coordination desk
Figure 3: Web work is safer when creative, design, deployment, and coordination surfaces stay distinct.
Conclusion: the winning workflow is a stack
The 2026 AI design stack is not collapsing into one winner. It is specializing. Claude Design explores experience and visual direction. Figma governs design systems and responsive collaboration. Lovart creates campaign visuals and brand-world variations. Sites in Codex turns compatible ideas and projects into hosted web experiences. MCPlato keeps research, source materials, approvals, iterations, and handoff observable.
The best teams will assign ownership: explore in Claude Design, govern in Figma, generate campaign assets in Lovart, turn suitable projects into hosted demos with Sites in Codex, and coordinate long-running work where evidence and decisions remain visible.