MCP 服务器

mockit-mcp

Turns text prompts into premium iOS UI mockups (PNG + HTML). Powered by Claude + a Playwright headless renderer. Two backends: local claude CLI (uses your Claude Code subscription) or the Anthropic API.

README

<h1>mockit-mcp</h1>

<p><strong>Turn text prompts into premium iOS mobile UI mockups.</strong></p> <p>An MCP server that pairs Claude (Opus 4.7 by default) with a Playwright renderer to generate <em>screenshot-grade</em> mobile app designs from natural language.</p>

What it does

Ask Claude Code (or any MCP client):

Design the home dashboard for a fitness tracker. Three concentric activity rings, weekly bar chart, recent workouts list, premium dark mode with neon accents.

mockit-mcp returns a real PNG mockup (sized 390×844 at 2x device scale, an iPhone-class viewport) and the underlying HTML/Tailwind source — so you can iterate visually and port to SwiftUI when you're ready to build.

It's not a static template engine and it's not generic AI slop. The system prompt is hand-tuned for premium iOS aesthetics: real content, SVG icons (no emoji), tasteful gradients in place of stock photos, iOS HIG type scale, and tonal layering instead of heavy shadows.

<div align="center"> <table> <tr> <td align="center"><img src="examples/screens/fitness.png" width="280" alt="Fitness tracker dashboard"/><br/><sub><b>Fitness dashboard</b></sub></td> <td align="center"><img src="examples/screens/volumetrik.png" width="280" alt="Volume calculator app screen (Turkish)"/><br/><sub><b>Volume calculator</b> (Turkish)</sub></td> </tr> </table> <p><sub>Generated from a single prompt each. See <a href="examples/">examples/</a>.</sub></p> </div>

Highlights

Two backends, same tools. Use the local claude CLI (subscription, $0 extra) or the Anthropic API (key + per-call pricing). Switch with one env var.
Real PNG output. Headless Chromium via Playwright. Default viewport is 390×844 @2x (iPhone-class); any custom size is one env var away.
Iterative refinement. iterate_screen takes a screen ID + feedback ("make the hero card smaller") and produces a new version, tracking parent/child.
Disk-backed library. Every generation saves HTML + PNG + JSON metadata. Browse, filter, re-export.
MCP standard. Works with Claude Code, Claude Desktop, Cursor, Windsurf, or any MCP client.
Stdio + HTTP transports. Run locally for dev, or as a network service for shared / containerized use.

Tools

Tool	Description
`generate_screen`	Text brief → PNG + HTML. Optional `design_system` and `project` fields.
`iterate_screen`	Take a previous `screen_id` + `feedback` string, produce a new version.
`list_screens`	List screens, optionally filtered by project.
`get_screen`	Fetch metadata (or full HTML) for a specific screen.

Install

Prerequisites

Node.js 20+
Either the claude CLI logged in (cli backend, default) or an Anthropic API key (api backend)
Playwright's Chromium download (~170 MB, one-time)

Quick start (CLI backend, recommended for local dev)

git clone https://github.com/karyaboyraz/mockit-mcp.git
cd mockit-mcp
npm install
npx playwright install chromium
npm run build

Add to Claude Code:

claude mcp add mockit -- node "$(pwd)/dist/server.js"

Done. No API key needed — it uses your existing claude CLI session.

API backend (no `claude` CLI on host)

echo "CLAUDE_BACKEND=api" > .env
echo "ANTHROPIC_API_KEY=sk-ant-..." >> .env
npm run build
claude mcp add mockit -- node "$(pwd)/dist/server.js"

Docker (HTTP transport, for shared deployment)

cat > .env <<'ENV'
CLAUDE_BACKEND=api
ANTHROPIC_API_KEY=sk-ant-...
# Required if you change the port binding from 127.0.0.1 to 0.0.0.0:
MCP_HTTP_TOKEN=$(openssl rand -hex 32)
ENV
docker compose up -d --build

By default docker-compose.yml binds the HTTP port to 127.0.0.1 only and the server requires MCP_HTTP_TOKEN for any non-loopback request. Don't expose this server to a public network without setting a strong MCP_HTTP_TOKEN — every generation hits your Anthropic API key.

Then point any client at the loopback URL:

claude mcp add --transport http mockit http://127.0.0.1:7821/mcp \
  -H "Authorization: Bearer <MCP_HTTP_TOKEN>"

For remote access, change docker-compose.yml's port binding to 0.0.0.0:7821:7821 and ensure MCP_HTTP_TOKEN is set — the server refuses to start otherwise.

Usage

In any MCP client, just ask:

Design a fitness tracker dashboard. Show today's ring progress, a weekly chart, and a list of recent workouts. Dark mode, neon green accent.

The PNG appears inline. The HTML is saved to designs/{project}/{name}-{id}.html.

For follow-ups:

iterate_screen on that fitness dashboard — replace the chart with heart-rate over time, and add a "share workout" button below.

See examples/ for prompt patterns and full outputs.

Configuration

All optional. See .env.example for the full list.

Env	Default	Notes
`CLAUDE_BACKEND`	`cli`	`cli` uses the `claude` CLI; `api` uses Anthropic SDK directly
`ANTHROPIC_API_KEY`	—	Required only for `api` backend
`ANTHROPIC_MODEL`	`claude-opus-4-7`	API backend only. If your account doesn't have Opus access, set to `claude-sonnet-4-6` or `claude-haiku-4-5`
`CLAUDE_CLI_PATH`	`claude`	Path to the `claude` binary
`CLAUDE_CLI_TIMEOUT_MS`	`180000`	Subprocess timeout
`MCP_TRANSPORT`	`stdio`	`stdio` or `http`
`HTTP_PORT`	`7821`	HTTP transport port
`HTTP_HOST`	`127.0.0.1`	Bind interface; non-loopback requires `MCP_HTTP_TOKEN`
`MCP_HTTP_TOKEN`	—	Bearer token for HTTP auth. Required if `HTTP_HOST` is non-loopback
`DESIGNS_DIR`	`./designs`	Where outputs are persisted
`VIEWPORT_WIDTH`	`390`	Render width in CSS pixels
`VIEWPORT_HEIGHT`	`844`	Render height in CSS pixels
`DEVICE_SCALE`	`2`	Retina factor (final PNG is `WIDTH × DEVICE_SCALE` wide)
`PLAYWRIGHT_NO_SANDBOX`	`auto`	`auto` = sandbox enabled outside containers; `true` to force-disable, `false` to force-enable. Disabling reduces isolation against malicious model HTML — only do so inside a container.

Cost

Per generation: ~3K input tokens (system prompt) + ~6–12K output tokens depending on screen complexity. Output dominates the cost on Opus.

Backend	First call	Cached follow-up
`cli`	counts against your Claude Code subscription quota	same — cache only discounts the system prompt
`api`	~$0.50–0.95 (Opus 4.7)	~$0.45–0.90 (cache discounts the system-prompt input only; output cost is unchanged)

System-prompt caching is on by default (5-minute TTL). It saves a few cents per call but is not an order-of-magnitude discount — output tokens still bill at full rate. For real cost reduction, switch to a smaller model (claude-sonnet-4-6 or claude-haiku-4-5).

Architecture

┌─────────────────┐
│   MCP Client    │  (Claude Code, Cursor, Windsurf, …)
└────────┬────────┘
         │ tool call: generate_screen({ prompt, ... })
         ▼
┌─────────────────────────────────────────────────────┐
│  mockit-mcp                                          │
│                                                      │
│  ┌──────────────┐    ┌────────────────────────┐    │
│  │  Backend     │    │  Renderer              │    │
│  │              │    │                        │    │
│  │ ► cli   ─────┼──► │  Playwright (headless  │    │
│  │ ► api   ─────┘    │  Chromium @ iPhone     │    │
│  │  → HTML+Tailwind  │  viewport)             │    │
│  └──────────────┘    │  → PNG screenshot      │    │
│                      └────────────────────────┘    │
│                                                      │
│  ┌────────────────────────────────────────────┐    │
│  │ Storage (disk): HTML + PNG + JSON metadata │    │
│  └────────────────────────────────────────────┘    │
└─────────────────────────────────────────────────────┘

Storage layout

designs/
└── {project-slug}/
    ├── {name-slug}-{id8}.html   # id8 = first 8 chars of the screen UUID
    ├── {name-slug}-{id8}.png
    └── {name-slug}-{id8}.json   # full UUID, prompt, parent ID, tokens, model, cost

Documentation

Reference and guides live in docs/:

Topic	Page
Install, first generation, verify	Getting Started
Every env var, defaults, what they do	Configuration
The four MCP tools — schemas, examples	Tools Reference
Pipeline end-to-end	Architecture
Local, Docker, networked deployments	Deployment
Sandbox, network allowlist, HTTP auth	Security
How the design voice is encoded	System Prompt
Common errors and how to fix them	Troubleshooting

Tuning the design voice

The hand-tuned system prompt lives in src/system-prompt.ts. It's where the iOS HIG enforcement, the no-stock-photo rule, the SF Pro fallback chain, and the editorial typography preferences are encoded. Want Material You instead, or a desktop dashboard voice? Edit it.

Development

npm run dev    # tsx watch mode, stdio transport
npm run http   # tsx watch mode, http transport on :7821
npm run build  # compile to dist/

Roadmap

[ ] Watch / iPad / Android viewport presets
[ ] Multi-screen flow generation (onboarding sequences)
[ ] HTML → SwiftUI / Jetpack Compose port tool
[ ] Design system import (Tailwind config, design tokens)
[ ] Image references (use --image for visual inspiration)
[ ] Variant generation (3-5 alternatives per prompt)

Contributing

Issues and PRs welcome — see CONTRIBUTING.md.

License

MIT

Acknowledgements

Built on top of:

Anthropic Claude — the model that does the heavy lifting
Model Context Protocol — the integration standard
Playwright — the renderer
Tailwind CSS — via CDN, in every generated screen

Trademarks

iPhone, iPad, Apple Watch, and iOS are trademarks of Apple Inc. Claude is a trademark of Anthropic, PBC. mockit-mcp is an independent open-source project and is not affiliated with, endorsed by, or sponsored by Apple Inc. or Anthropic, PBC. All other product names, logos, and brands are property of their respective owners.