DesignDiff MCP
Verifies that AI-generated UI code matches Figma design specs by rendering components in a real browser, comparing computed CSS, and returning patch-ready fixes with scored parity reports.
README
DesignDiff MCP
AI implementation verification engine for Figma and MCP-compatible coding agents.
DesignDiff verifies that AI-generated interfaces match your Figma design system before they ship. It reads the Figma spec, renders your component in a real browser, compares computed CSS output, and returns patch-ready fixes — all inside your existing coding agent session.
Claude Code / Cursor / Windsurf
↓
DesignDiff MCP
↓
Figma REST API + Playwright renderer
↓
Scored diff + Patch-ready fix
↓
Agent applies → Re-checks: 94/100
Why
AI agents generate UI code from Figma specs and get it mostly right — but not exactly. The typical failure pattern:
- Token violations — hardcoded
#2563EBinstead ofvar(--brand-blue-600). Works today, breaks on the next brand refresh. - Spacing drift — padding is 12px not 16px because the agent approximated instead of reading the exact value.
- Missing states — hover, focus, and disabled were defined in Figma variants the agent didn't parse.
The component looks correct visually. That's what makes this dangerous. DesignDiff catches these errors before they land in a PR.
Quick start
1. Get your Figma API key Settings → Account → Personal access tokens
2. Add to your agent config
Claude Code (~/.config/claude/mcp.json):
{
"mcpServers": {
"designdiff": {
"command": "npx",
"args": ["-y", "designdiff-mcp"],
"env": { "FIGMA_API_KEY": "your-token-here" }
}
}
}
Same format for Cursor (.cursor/mcp.json), Windsurf, VS Code Copilot, and Zed.
3. Say to your agent
"Build the Button component from Figma file abc123, node 123:456,
running at localhost:6006 — and verify it matches the spec"
DesignDiff is called automatically when the agent generates or modifies a component.
Works best with Storybook, local preview routes, or any stable component URL your agent can render.
Tools
check_component_parity — ships now
Diffs the Figma design spec against computed browser CSS. Returns a scored parity report with ranked mismatches and a patch-ready fix when score drops below threshold.
check_component_parity({
file_id: "abc123", // Figma file ID from URL
node_id: "123:456", // Component node ID
component_url: "http://localhost:6006/story/button--primary",
code_path: "src/components/Button.tsx",
threshold: 80, // Score below which fix is generated (default: 80)
})
What it checks: spacing, color tokens, typography, borders, border-radius, interactive states.
What it returns: 0–100 parity score, ranked mismatches with consequences, pattern detection (e.g. "AI-generated code signature"), and a patch-ready git diff.
flag_stale_mappings — ships now
Cross-references Code Connect mappings in your repo against Figma version history and git blame. Surfaces components that have drifted silently, ranked by blast radius and days since divergence.
flag_stale_mappings({
file_id: "abc123",
repo_root: ".", // Path to your repo root
})
Reads from .figma/code-connect.json in your repo root.
generate_sync_patch — ships now
Generates a surgical patch for a specific component — token substitutions, prop corrections, missing state stubs. Output as git diff, JSX corrections, or CSS delta.
generate_sync_patch({
file_id: "abc123",
node_id: "123:456",
code_path: "src/components/Button.tsx",
format: "diff", // "diff" | "jsx" | "css"
})
audit_state_coverage — v0.2
Checks whether every interactive state defined in a Figma component set (hover, focus, disabled, error, loading) exists in code. Flags WCAG 2.4.7 risk for missing focus rings.
check_responsive_parity — v0.2
Renders at 375/768/1280/1920px and catches overflow, collapsed containers, and flex direction errors.
check_theme_parity — v0.2
Runs parity across Figma variable modes — light/dark, brand variants, high-contrast. Catches hardcoded colors that are invisible in light mode but break in dark.
Finding your Figma IDs
File ID: From the URL — figma.com/file/{FILE_ID}/...
Node ID: Right-click a component in Figma → Copy link → extract the node-id parameter. Format: 123:456. Note: the URL uses 123-456 (hyphen) but the API uses 123:456 (colon) — DesignDiff handles the encoding automatically.
How it works
-
Fetch — Calls the Figma REST API and extracts the full design spec: exact pixel values, token references, and every variant state from the component set.
-
Render — Launches a headless Playwright browser and renders your component at its real URL. Captures computed CSS — what the browser actually produces after cascade and runtime overrides resolve.
-
Score — Compares the two. Token violations score highest (they break the entire design system contract). Missing focus states are flagged with WCAG reference. Every mismatch includes the real-world consequence of not fixing it.
-
Fix — When score is below threshold, generates a patch-ready diff the agent applies in the same response, then re-verifies the score improved.
Configuration
| Environment variable | Required | Default | Description |
|---|---|---|---|
FIGMA_API_KEY |
✓ | — | Figma personal access token |
TRANSPORT |
— | stdio |
stdio or http |
PORT |
— | 3847 |
HTTP port (when TRANSPORT=http) |
HTTP transport
For use with non-stdio MCP clients:
FIGMA_API_KEY=xxx TRANSPORT=http npx designdiff-mcp
# → DesignDiff MCP v0.2 running at http://localhost:3847/mcp
Health check: GET /health
Development
git clone https://github.com/designdiff/designdiff-mcp
cd designdiff-mcp
npm install
npx playwright install chromium
npm run build
FIGMA_API_KEY=xxx npm start
Inspect tools interactively:
npm run inspector
This opens the MCP Inspector UI where you can call tools directly.
Roadmap
| Stage | What ships |
|---|---|
| v0.1 — Now | Rendered parity check · token compliance · patch-ready diffs |
| v0.2 | State coverage · responsive parity · theme parity · Storybook adapter |
| v1.0 | Unified quality score · CI gate · score history · Slack alerts |
Contributing
See CONTRIBUTING.md.
License
MIT — see LICENSE.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。