A developer building a landing page for a new conversational developer-agent product started the way many AI-assisted design workflows start today: Claude Code connected to Figma through the Figma Model Context Protocol (MCP) server, the standard interface that lets an AI agent read and write a tool's data. After a handful of design prompts, the Figma MCP monthly rate limit kicked in and blocked further work for the rest of the billing month. The detour that followed led to Claude Design, Anthropic's design environment inside the Claude ecosystem, and ended with a workflow where the design canvas, the comment thread and the coding agent all operate on the same underlying code. The result, in this case, was a full replacement of Figma for web interface work.
Why the Figma MCP route stalled
The first attempt did produce a landing page in Figma. The layout and typography looked well structured, but everything on the artboard was static: no working buttons, no hover effects, no component states. Generation through the MCP also produced small, recurring defects such as missing SVG icons, misaligned glyphs and a broken copy-to-clipboard icon, each of which needed a corrective prompt. Those corrections consumed tool calls, and the tool calls are what Figma meters.
Figma's own documentation lists the quotas for its MCP server, which it markets to clients including Claude, Claude Code, ChatGPT, Codex by OpenAI, Cursor, Xcode and VS Code. The limits vary by plan and by seat type:
| Plan | View and Collab seats | Full Dev seats |
|---|---|---|
| Starter (free) | Up to 20 tool calls per month | — |
| Professional | 20 calls per month | 200 per day, burst of 10 per minute |
| Organization | 20 calls per month | 200 per day, burst of 15 per minute |
| Enterprise | 20 calls per month | 600 per day, burst of 20 per minute |
Seat prices on the Professional tier are $20 per month for a Full seat ($16 per month billed annually), $15 per month for a Dev seat ($12 annually) and $5 per month for a Collab seat ($3 annually). Organization full seats cost $55 per month and Enterprise full seats $90 per month.
Read against an agent workflow, these numbers appear hard to reconcile with real use. For an iterative agent session, 20 calls a month looks closer to a trial than a working allowance, and even the daily Full Dev quotas seem tight for an agent that may fire several tool calls per prompt. The argument here is that paying customers, especially on enterprise plans, should not face per-minute and per-day throttles that break agent loops. Whether Figma changes these limits is an open question; for now they are the published terms.
What Claude Design produces instead
Claude Design, first introduced by Anthropic Labs on April 17, 2026, is included in standard Claude subscriptions. There is no third-party connector to configure and no separate per-call quota for design work. A single prompt issued from the Claude Code terminal generated the same landing page as a Claude Artifact, Claude's format for self-contained, runnable outputs, and opened it in the Claude Design editor.
The difference from the Figma output is that the artifact runs. In full-screen presentation mode, the page is a live web application: JavaScript executes inside the canvas, hover transitions fire, custom animations play and components hold state. The page included an embedded agent console widget that streams simulated log traces, animates step verification and responds to clicks on buttons such as "Yes, open it" and a pull-request action. A sample scenario inside the widget investigates a question like why p95 latency spiked on a checkout flow overnight, pulling metrics and logs before proposing a pull request.
Claude Code also appeared to hit the intended design more accurately when building natively in Claude Design than when going through the Figma MCP, with fewer of the small glyph and icon defects that marked the Figma run. Some "vibe-coded" aesthetic quirks remain in generated designs, but the working interactivity seems to outweigh them.
The inspector looks familiar
Figma users will recognize the right-hand inspector. Selecting any element on the canvas, such as a heading, exposes its design and CSS properties for direct editing:
- Typography: font family (Geist, for example), weight, size expressed as CSS clamp values such as clamp(56px, 8.5vw, 144px), line height, letter spacing, alignment and hex color
- Layout: per-side margins, padding, border radius, fill colors and alignment toggles
- Appearance and Effects: drop shadows, layer blur, background blur, transforms, opacity and custom borders
- Advanced: a debugging and styling section
- Export: elements or whole frames to PNG and other formats at multiple scales
When Claude Design first shipped it felt basic, but the recent updates have brought it close to Figma's core editing set, and the canvas supports the zooming and panning across multi-page mockups that designers expect.

▲ Multi-page prototype with linked navigation
Multi-page prototypes with real navigation
Prototyping, one of Figma's most-used features, works inside Claude Design as well. The developer-agent prototype links a home page with Pricing, Developers and Capabilities pages, and navigation links jump to specific sections. Because the pages are code, the transitions behave like a production site rather than a click-through of flat frames.
The pricing page was generated with a follow-up instruction to add a matching view inside the same artifact file and wire the navigation together. It came back with a monthly and yearly billing toggle, side-by-side plan comparisons, feature checklists and animated cards for Hobby, Pro and Scale tiers at $0, $49 and $299 per month, with a 14-day trial on the paid cards and call-to-action buttons that route back to the home hero. These are placeholder prices generated for the prototype, not a real product's pricing.
One snag surfaced during iteration: the system accessibility setting prefers-reduced-motion froze the hero animation. Claude adjusted the script to keep subtle movement active. If an artifact animation stops in preview, reduced-motion queries and visibility observers are a sensible first place to look.
Comment on the canvas, Claude Code edits the code
The feature that most changes the day-to-day loop is in-canvas commenting. The usual way to ask an agent for a visual change is to take a screenshot, paste it into chat and describe where the element sits. In Claude Design, a comment pin attaches to the exact element.
In one test, a pin on the "Docs" link in the navigation bar carried the instruction "let's remove this docs link." The Claude Code session running in the background detected the new artifact comment thread on its own, printed a notification that it was processing the thread, located the anchor element around line 532 of the artifact's HTML file, removed the link and marked the thread resolved. Reloading the artifact showed the link gone. No command was typed in the terminal. The same mechanism works for a request such as shrinking a hero heading, tweaking copy or adjusting layout.
The prompt behind the design
The landing page came from a deliberately specific prompt to Claude Code, which was running Claude Opus 3.5 at medium effort. Its main components:
- A persona: a world-class creative web designer specializing in premium AI and developer tools.
- An explicit instruction to use Claude's native design capability and Claude Artifacts rather than external tools. This line matters; without it, Claude Code may default to a configured Figma MCP integration.
- Reference sites. modal.com supplied the dark-mode typography, high-end feel and glowing ASCII-style 3D particle hero. firecrawl.dev supplied the developer-centric styling, subtle animated background grid, interactive command boxes and syntax-highlighted code blocks.
The output reflected both: a rotating glyph-particle sphere in the hero, an interactive terminal demo with typewriter-style responses and agent traces in the center, a "$ pip install" snippet with a copy button, and a developer section headed "Fifteen lines to production." Asking the model to synthesize patterns from two or three distinct, well-regarded sites seems to produce a more distinctive result than a generic "make it look modern" instruction.

▲ Artifact converted into a running React app
From artifact to a running React app
The reason the artifact-to-code step works without a translation layer is that Claude Design and Claude Code share the same code representation. The design is not a picture of a page that must be re-described to a coder; it is already markup, styles and scripts the agent understands.
A single command asked Claude Code to build a runnable app matching the artifact using Vite, React, TypeScript and Tailwind CSS. Claude Code scaffolded the project, installed dependencies, mapped each UI element into React components and started a local Vite dev server at localhost:5173. The page in the browser matched the artifact: working tabs, the typewriter console animation, routing between pages, CSS transitions, the dark theme and glow effects, with no broken component. What used to be days of mockup iteration and a separate handoff to frontend code collapsed into a single afternoon, according to this experience.
Where Figma still fits
This is not a claim that Figma has no role. For graphic-heavy work such as custom vector illustration and video thumbnails, where fine manual control of the canvas matters, Figma remains the better tool in this workflow. The replacement is specific to web interfaces, landing pages and interactive software products, where a static artboard is a weaker artifact than a runnable one.
The broader interpretation is that static interface mockups are being superseded by interactive, code-backed artifacts, and that standalone design tools will need agent-friendly integrations without tight call quotas to stay in that loop. That is a forecast, not a settled outcome.
What to take from this
- If you rely on Figma MCP for iterative agent work, check your seat type against the quota table above before you plan a session; collaborator seats get 20 calls per month.
- Claude Design is already part of a Claude subscription, so trying it costs nothing extra.
- In prompts, state explicitly that Claude Code should use its native design capability and Claude Artifacts, and supply two or three live reference URLs for typography, layout and animation style.
- Use full-screen presentation mode to test clicks, hover states and transitions on the canvas itself.
- Leave comment pins on specific elements instead of describing coordinates in chat; Claude Code picks them up in the background.
- Use the inspector for clamp font sizes, padding, hex colors and export scales rather than editing code by hand.
- When the design is final, ask Claude Code to scaffold a Vite, React, TypeScript and Tailwind CSS project from the artifact and verify it on the local dev server.


