Figma Live Design QA MCP
Compares Figma frames to live pages, checking colors, fonts, and border radii, and generates a shareable HTML report.
README
Figma Live Design QA MCP
Figma Live Design QA MCP helps you answer a simple question:
Does the page we built still look and feel like the design?
Give it a Figma frame and a live page. It checks the design details that should match — colors, type, fonts, and corner radii — then gives you a shareable HTML report.
It is made for designers, design engineers, and frontend teams who want a quick way to catch “almost right” details before handoff or release. You do not need to be a developer to use the actual QA workflow. A little setup is needed the first time.
What it checks
- Colors and near-match colors
- Font family loading
- Font sizes and weights
- Border radii, including pills
- Repeated issues, with useful locations in the page
- A report with design and build screenshots where available
It does not do a pixel-perfect screenshot diff. That is intentional: Figma often uses placeholder content while the real page uses live data, so pixel diffs can create a lot of noise.
Setup
You only need to do this once.
1. Install the project
You need Node.js 20 or newer. From this folder, run:
npm install
npx playwright install chromium
npm run build
2. Create a Figma token
Create a Figma personal access token with File content: read permission. You will add it to Claude’s configuration in the next step.
Keep it private: do not paste it into a Figma URL, a prompt, a screenshot, or a GitHub file. See Security.
3. Add the server to Claude Desktop
-
Open Claude Desktop.
-
Go to Settings → Developer → Edit Config.
-
Add this inside the
mcpServersobject. Replace the example path with the real path to this project:{ "mcpServers": { "figma-live-design-qa": { "command": "node", "args": ["/absolute/path/to/figma-live-design-qa-mcp/dist/index.js"], "env": { "FIGMA_TOKEN": "figd_your_token_here" } } } } -
Save the file and restart Claude Desktop.
-
Check that Claude can see the Figma Live Design QA tools.
4. If your site needs a login
Start Chrome with remote debugging, sign in to your staging site in that window, and leave it open:
/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome --remote-debugging-port=9222
The server reuses that signed-in session. It does not ask for or store your password. More setup details are in Setup.
How to use it in Claude
- In Figma, select the frame you want to check and choose Copy link to selection.
- Open the matching page in your local, preview, or staging build.
- Paste both links into Claude and ask for a design QA check.
- Tell Claude about anything important: viewport, theme, locale, or intentional exceptions.
- Open the HTML report from
~/Documents/Design QA/.
Sample prompt
Run design QA for this checkout page.
Figma frame: https://www.figma.com/design/FILE_KEY/Checkout?node-id=12-34
Live page: https://staging.example.com/checkout
Viewport: use the Figma frame width
Theme: light
Locale: en-US
Compare the colors, font family, font size, font weight, and border radius. Highlight
near-token differences as likely hardcoded drift, explain what looks intentional, and
generate the self-contained HTML report. If a dimension cannot be checked confidently,
say that clearly instead of guessing.
Short version:
Run design QA on
https://www.figma.com/design/FILE_KEY/File?node-id=12-34againsthttps://staging.example.com/checkout.
If one stage has a problem, Claude gets a runId so it can retry that stage instead of starting over.
Make it fit your design system
The quality of the result depends on the frame you choose. Think of the Figma frame as the reference board for the check.
- Pick a frame that shows the colors, type styles, and radii you care about.
- Make sure important values are connected to Figma variables or published styles when possible.
- Tell Claude which theme, breakpoint, locale, and font-loading state to use.
- Call out things that are intentionally different, like third-party widgets or bespoke marketing artwork.
- If a dimension says
not_verified, it means there was not enough design evidence to make a fair call — not that the page failed.
The check currently covers colors, font loading, font sizes, font weights, and border radii. It does not check copy, layout geometry, line-height, or pseudo-element styles.
Share the result
The report is a self-contained HTML file you can open without the app running. If you use Claude Cowork for review and handoff, ask it to read the report, summarize the biggest issues, and export the review as HTML, DOC/DOCX, or PDF.
For example:
Review the design QA report. Group the findings by severity and design-system dimension, keep the original values and occurrence counts, add a short summary for designers and product, and export the review as an HTML file and a PDF. Do not include API tokens, cookies, credentials, or private configuration values.
Please review reports before sharing. They can contain private product copy, URLs, and screenshots from the build.
The tools, in plain English
| Tool | What it does |
|---|---|
run_design_qa |
Does the whole check in one go. This is the usual starting point. |
capture_design |
Reads the selected Figma frame. |
capture_build |
Reads the live page in your browser session. |
check_tokens |
Compares the page values with the design values. |
build_report |
Creates the HTML report and JSON details. |
dismiss_finding |
Remembers a known, intentional exception for that Figma file. |
Where files go
- Reports:
~/Documents/Design QA/ - Run details:
~/.figma-live-design-qa-mcp/runs/ - Dismissed findings:
~/.figma-live-design-qa-mcp/baselines/
You can change these locations with DESIGN_QA_HOME and DESIGN_QA_REPORT_DIR.
For the technical folks
npm test
npm run typecheck
npm run build
The project is a local stdio MCP server. It uses Playwright to read the live page, stores detailed run data locally, and keeps the model response compact so large DOM trees do not flood the conversation.
More detail:
A few things to know
- Figma and the live page can use different content and still pass the useful parts of the check.
- Cross-origin iframe contents are skipped.
- A light/dark mode mismatch can make color comparison
not_verified. - The Variables REST endpoint is Enterprise-only; the server also works from bound values and repeated values in other plans.
License
MIT License
Copyright (c) 2026 Madhav M Nair
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
IN NO EVENT SHALL THE AUTHOR OR COPYRIGHT HOLDER BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
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。