figcraft
Enables AI-powered design operations in Figma, including UI creation, design review, token sync, linting, auditing, and auto-fix through natural language commands.
README
FigCraft
English | 中文
AI-powered Figma plugin for design quality. Two-way bridge between AI IDEs and Figma — create UI, review designs, sync tokens, lint for compliance, audit, and auto-fix, all via natural language. Works great on its own, and even better alongside the official Figma MCP server.
New here? Start with docs/introduction.md for a 5-minute product tour (positioning, three core capabilities, FAQ). This README focuses on installation and reference.
What can you do with it?
Describe what you want in natural language, and FigCraft + Figma MCP make it happen in Figma:
"Create a login screen, then lint the whole page and auto-fix issues"
"Sync tokens from my DTCG JSON to Figma variables, diff and update"
"Check WCAG contrast and target sizes on this page, auto-fix what you can"
Features
- 🎨 From creation to delivery, fully covered — Create UI directly in Figma with 116 MCP tools. Build frames, components, variants, icons — check quality right after, fix issues on the spot
- 🧠 Opinion Engine — auto-infers layout direction, sizing, token bindings, and catches parameter conflicts before they hit Figma. You describe what, it figures out how
- 🔍 Automated design audit — token bindings, color contrast, spacing, component health — all checked in one pass · Full guide →
- 🔧 Lint + fix in one step — 40 rules covering token compliance, WCAG, layout structure — one command to batch-fix everything flagged
- 🔄 Two-way token sync — DTCG JSON ↔ Figma variables, Light/Dark multi-mode in one step. Changed tokens in code? Just sync
- 🔀 Dual mode for any team — Library mode for Figma shared libraries, Spec mode for DTCG JSON — pick what fits your workflow
- 📐 Prototype → dev docs — parse prototype interactions into Mermaid flow diagrams + interaction specs, no more manual handoff docs
- 🛡️ Harness Pipeline — auto-verifies every creation, recovers from errors with actionable suggestions, and tracks quality debt across turns
Quick Start
Requires Node.js >= 20.
1. Install the Figma Plugin
FigCraft is not yet on the Figma Community. Build from source:
git clone https://github.com/DivikWu/figcraft.git
cd figcraft
npm install
npm run build
Then in Figma Desktop:
- Plugins → Development → Import plugin from manifest
- Select the
manifest.jsonfile from the cloned repo
2. Add MCP Servers to your IDE
FigCraft handles both UI creation and design quality on its own. For even more creation capabilities, you can add the official Figma MCP server alongside it — both servers run in parallel and complement each other.
Note:
figcraft-designis not yet published to npm. You need to build from source first (step 1 above). Replacecwdbelow with the absolute path to your local clone.
FigCraft config (same for all IDEs):
{
"mcpServers": {
"figcraft": {
"command": "node",
"args": ["dist/mcp-server/index.js"],
"cwd": "/your/absolute/path/to/figcraft"
}
}
}
FigCraft works standalone for both UI creation and design quality. Adding the Figma MCP server gives you even more creation capabilities.
<details> <summary><strong>Adding the official Figma MCP server (for extra creation capabilities)</strong></summary>
Figma provides two deployment options:
Desktop server (local, runs inside Figma Desktop App):
- Open Figma Desktop → Dev Mode → Enable MCP server in the inspect panel
- Add to your IDE config:
{
"mcpServers": {
"figma-desktop": {
"url": "http://127.0.0.1:3845/mcp"
}
}
}
Remote server (cloud, broader feature set — recommended by Figma): See Figma's remote server setup guide.
For full details, see the official Figma MCP documentation. </details>
Put it in the right file for your IDE:
📦 If you cloned this repo for fork-and-use, the shared project configs are already committed:
.cursor/mcp.json,.mcp.json,.kiro/settings/mcp.json.example— plus.claude/settings.jsonfor Claude Code's auto-approval. Copy the Kiro example to.kiro/settings/mcp.jsonfor local Kiro use. Each IDE has a different auto-approval mechanism — see the per-IDE notes below or user-guide §6.5 for the full table.
<details> <summary><strong>Cursor</strong> — <code>.cursor/mcp.json</code></summary>
Create .cursor/mcp.json in your project root with the config above.
Auto-approval is configured separately at user level — Cursor does NOT auto-approve from .cursor/mcp.json. Create ~/.cursor/permissions.json:
{ "mcpAllowlist": ["figcraft:*", "figma-desktop:*"] }
(See Cursor permissions docs. Auto-Run mode must be enabled in Cursor settings for the allowlist to apply.) </details>
<details> <summary><strong>Claude Code</strong> — <code>.mcp.json</code></summary>
Create .mcp.json in your project root with the config above.
Auto-approval: Claude Code ignores any autoApprove field inside .mcp.json (that field is a Cursor/Kiro extension, not MCP standard). Use .claude/settings.json instead:
{
"permissions": {
"allow": ["mcp__figcraft__create_frame", "mcp__figcraft__nodes", "..."]
}
}
This repo's pre-committed .claude/settings.json already lists all 122 figcraft approval entries (119 schema tools + 3 toolset meta tools) + 13 figma-desktop tools. Run npm run schema to regenerate after upstream tool changes.
</details>
<details> <summary><strong>Kiro</strong> — <code>.kiro/settings/mcp.json</code></summary>
Copy .kiro/settings/mcp.json.example to .kiro/settings/mcp.json in your project root. Kiro supports autoApprove natively:
{
"mcpServers": {
"figcraft": {
"command": "npx",
"args": ["tsx", "packages/figcraft-design/src/index.ts"],
"cwd": "${workspaceFolder}",
"disabled": false,
"autoApprove": ["ping", "create_frame", "nodes", "..."]
}
}
}
This repo's pre-committed .kiro/settings/mcp.json.example already auto-approves all 122 figcraft approval entries (119 schema tools + 3 toolset meta tools). Tools are surfaced in Kiro as mcp_figcraft_<tool> (e.g. mcp_figcraft_ping).
Tip: This repo includes
.kiro/steering/figcraft.mdas a workflow guide. Copy it to your project's.kiro/steering/folder. </details>
<details> <summary><strong>Antigravity (Google)</strong> — MCP Server management panel</summary>
Open Antigravity → Agent dropdown → Manage MCP Servers → View raw config, then paste the config above. </details>
<details> <summary><strong>Codex CLI (OpenAI)</strong> — <code>~/.codex/config.toml</code></summary>
[mcp_servers.figcraft]
command = "node"
args = ["dist/mcp-server/index.js"]
cwd = "/your/absolute/path/to/figcraft"
</details>
3. Connect & Verify
Open the FigCraft plugin in Figma — both sides auto-connect via the WebSocket relay. The plugin UI shows the channel ID and connection status.
To verify the connection works, ask your AI IDE to run the ping tool. If it returns a response, you're all set.
Troubleshooting: If the connection fails, check that port 3055 is not occupied by another process. The relay will auto-try ports 3056–3060 as fallback.
Architecture
FigCraft operates on a single Plugin Channel to Figma:
AI IDE (Kiro / Cursor / Claude Code / Antigravity / Codex)
│ MCP (stdio)
▼
MCP Server (Node.js)
└── Plugin Channel ──→ WS Relay (:3055) ──→ Figma Plugin
(lint, audit, token sync, node ops)
- Plugin Channel: WebSocket relay to the FigCraft Figma Plugin. Required for all operations — creation, lint, audit, node inspection, and token sync all run through the Plugin API sandbox.
pingchecks Plugin Channel connectivity and reports status.- FigCraft handles UI creation, design system search, component management, and design-to-code context extraction natively. The official Figma MCP server complements it with FigJam support and Code Connect publishing.
Dual Mode
| Mode | Token Source | Lint Behavior | Use Case |
|---|---|---|---|
| Library | Figma shared library | Check variable/style bindings | Daily design with team library |
| Spec | DTCG JSON files | Check values against token specs | Spec-driven validation |
Switch modes via set_mode tool or the plugin UI.
UI Creation
FigCraft creates UI directly in Figma — frames, text, SVG, components, variants, icons, and images. The Opinion Engine auto-infers layout, sizing, and token bindings so you describe structure, not implementation details. GRID layout, nested children trees, and batch operations are all supported.
create_framewith inlinechildrenbuilds entire screen hierarchies in one callcreate_component/create_component_setbuild reusable component libraries with variant guardrailscreate_image_frame_from_localimports local PNG/JPEG/GIF files as image-filled frames;fill_existing_image_from_localreplaces existing fills with edit accessimport_htmlconverts static HTML, local.htmlfiles, or simple URLs into editable FigCraft frame/text/image layers- After creating UI, the Harness Pipeline auto-verifies quality; or run
lint_fix_allmanually - Use
get_design_contextto extract node trees with resolved tokens for design-to-code workflows
Lint Rules (40)
Current lint coverage spans token compliance, WCAG accessibility, layout structure, screen-level quality, naming, and component health.
- Token compliance (5): color, typography, radius, hardcoded token usage, missing text style
- WCAG accessibility (5): contrast, target size, text size, line height, non-text contrast
- Layout & Structure (27): screen shell validation, misclassified interactive root, nested interactive shell, missing auto-layout, empty container, spacer frames, nesting depth, button variants (solid/outline/ghost/text/icon), standalone link, text overflow, form consistency, CTA width consistency, overflow parent, HUG/STRETCH paradox, section spacing collapse, screen bottom overflow, social/nav/stats row crowding, input field structure, mobile dimensions, elevation consistency, elevation hierarchy
- Naming & Content (2): default name detection, placeholder text detection
- Component (1): component property binding checks
See docs/generated/lint-rules.md for the complete rule reference.
Environment Variables
| Variable | Description | Default |
|---|---|---|
FIGCRAFT_RELAY_PORT |
Relay WebSocket port | 3055 |
FIGCRAFT_RELAY_URL |
Full WebSocket relay URL (overrides port) | ws://localhost:3055 |
FIGCRAFT_CHANNEL |
Channel ID | figcraft |
FIGMA_API_TOKEN |
Figma Personal Access Token (for REST API fallback; can also be set in plugin UI or via OAuth) | — |
FIGCRAFT_ACCESS |
Access control level: read, create, or edit |
edit |
Development
Requires Node.js >= 20.
npm install
npm run build # Build all (MCP server + relay + plugin)
npm run build:plugin # Build Figma plugin only
npm run dev:relay # Start WebSocket relay server (for debugging)
npm run dev:mcp # Start MCP server (stdio transport)
npm run schema # Regenerate tool registry from schema/tools.yaml
npm run content # Compile templates, guides, and prompts from content/
npm run typecheck # TypeScript type check
npm run test # Run unit tests (vitest)
For details on content assets (templates, guides, prompts) and how to add new ones, see docs/asset-maintenance.md.
<details> <summary><strong>Run MCP server from source (for development)</strong></summary>
Instead of npx figcraft-design, point your IDE to the local source:
{
"mcpServers": {
"figcraft": {
"command": "npx",
"args": ["tsx", "packages/figcraft-design/src/index.ts"],
"cwd": "/path/to/figcraft"
}
}
}
</details>
Contributing
Contributions welcome! Fork the repo and open a Pull Request.
Before submitting, make sure:
npm run typecheck # Type check passes
npm run test # Tests pass
License
MIT
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。
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。
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。
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。