Skip to main content
Glama

designer-mcp

Cursor-style designer pen for Claude Code. Click, marquee, or draw on any webpage in a headed Chromium and Claude gets back the exact source file, line number, CSS selector, and a screenshot — ready to edit and verify.

What it does

Three modes of visual-to-source:

Mode

Interaction

Claude gets back

element

Hover + click one element

{ selector, tag, classes, text, html, rect, source: { fileName, lineNumber, componentName }, screenshot_path }

area

Drag a marquee

{ rect, elements: [{ selector, source, rect, ... }], screenshot_path }

draw

Freeform red pen, Enter to finish

{ strokes, viewport, screenshot_path (pen only), viewport_screenshot_path (page + ink) }

All screenshots are saved as PNG files in /tmp and returned as paths — your MCP client never hits a context limit on base64.

React source resolution works in Next.js dev mode via the _debugSource fiber property (attached by @babel/plugin-transform-react-jsx-source). Production builds strip this; see Production source mapping below.

Related MCP server: Browser Feedback MCP

Demo

You:     "Make this button rounder"
Claude:  [designer_open http://localhost:3000/dashboard]
Claude:  [designer_pick mode=element]
You:     *click the button*
Claude:  → source: Button.tsx:42
Claude:  [Edit Button.tsx add rounded-full]
Claude:  [designer_screenshot selector=#cta-btn]  ← after screenshot for verification

Install

Prereqs: Node 18+, Claude Code, a working macOS/Linux (Playwright Chromium).

git clone https://github.com/YOUR_USER/designer-mcp.git
cd designer-mcp
npm install
npx playwright install chromium      # one-time browser download

Register the MCP with Claude Code (user-scope = available in every session):

claude mcp add --scope user designer-mcp node "$(pwd)/index.js"

Install the Claude skill so future sessions know the workflow:

mkdir -p ~/.claude/skills/designer
cp SKILL.md ~/.claude/skills/designer/SKILL.md

Restart Claude Code. You should see designer_* tools and a designer: skill in your session.

Usage

Start your Next.js dev server (for source mapping):

cd your-nextjs-app && npm run dev

Then, in Claude Code:

"Open http://localhost:3000/settings in the designer and let me pick the header."

Claude will call designer_open(...), then designer_pick({ mode: "element" }). Chromium pops to the front, your cursor becomes crosshairs, you click the header. Claude gets source.fileName + lineNumber and can edit directly.

Modes cheat sheet

  • Single element — use element

  • Multiple related elements in one region — use area (drag a box; returns every element whose center falls inside)

  • Annotate / explain visually — use draw (red pen, Enter to finish, Esc to cancel)

Production source mapping

_debugSource is dev-only. To use the picker on a production build, enable source maps in next.config.js:

module.exports = {
  productionBrowserSourceMaps: true,
  // ...
};

The picker currently returns source: null in prod; a future version will resolve the selector through the deployed sourcemap. PRs welcome.

Tool reference

All tools are exposed over MCP; Claude Code sees them as mcp__designer-mcp__*.

designer_open(url: string)

Launch or reuse the headed Chromium instance and navigate. Foregrounds the window on macOS via bringToFront() + an AppleScript nudge.

designer_pick({ mode?: "element" | "area" | "draw" })

Activate the picker overlay. Returns when the user completes the interaction (or Esc cancels, or 180s timeout).

designer_screenshot({ selector?: string })

PNG of the page or a specific element. Returns { path, bytes }.

designer_close()

Tear down the browser and release Playwright resources.

How it works

  1. A Playwright-controlled Chromium is launched headed. Singleton per process.

  2. designer_pick injects a small vanilla-JS overlay (picker.js) into the page. The overlay:

    • element mode — tracks mousemove/click, outlines the hover target in blue, resolves a unique-ish CSS selector, walks the React fiber chain for _debugSource, returns to the MCP.

    • area mode — rubber-band marquee; on mouseup, any element whose center falls inside the box is collected (dedup by selector).

    • draw mode — full-viewport canvas overlay; captures strokes as point arrays; Enter finishes.

  3. The server polls window.__designerResult every 200ms for up to 180 seconds.

  4. On completion, an appropriate screenshot (element / area clip / full viewport) is saved to /tmp and the path returned.

Contributing

  • PRs welcome, especially around:

    • Production sourcemap resolution

    • Kestrel/React Native picker (currently web only)

    • Multi-element accumulation in element mode (Cmd-click to add)

    • VS Code "reveal in editor" integration

License

MIT

Available Tools

4 tools
designer_closeA

Close the designer browser and release resources.

Inputs: no named parameters.

designer_openA

Open a URL in the designer's headed Chromium (launches it if not running). Use this before designer_pick.

Inputs: url (required).

designer_pickA

Activate the picker in the designer browser. Three modes: element — user clicks one element; returns { selector, tag, classes, text, html, rect, source, screenshot_path } area — user drags a marquee; returns { rect, elements: [{selector, source, rect, ...}], screenshot_path } draw — user ink-annotates with a red pen, Enter to finish; returns { strokes, viewport, screenshot_path (strokes only), viewport_screenshot_path (full view with drawings) } Esc cancels in any mode. screenshot_path / viewport_screenshot_path point to PNGs in /tmp; open with the Read tool.

Inputs: mode.

designer_screenshotA

Screenshot the current page or a specific element selector. Returns { path, bytes } — a filesystem path to a PNG in /tmp that you can Read with the Read tool.

Inputs: selector.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.3/5.0

    Scored across 4 tools

    Disambiguation5/5

    Each tool has a clearly distinct purpose with no overlap: open launches the browser, pick activates the picker with specific modes, screenshot captures images, and close terminates the session. The descriptions clearly differentiate their functions, eliminating any ambiguity in tool selection.

    Naming Consistency5/5

    All tool names follow a consistent 'designer_' prefix with descriptive action suffixes (open, pick, screenshot, close), using snake_case uniformly. This predictable pattern makes the tool set easy to navigate and understand at a glance.

    Tool Count5/5

    With 4 tools, this server is well-scoped for its purpose of browser-based design interactions. Each tool earns its place by covering essential operations: launching, picking elements, capturing screenshots, and cleaning up, without being overly sparse or bloated.

    Completeness5/5

    The tool set provides complete lifecycle coverage for the designer domain: it supports opening the browser, interactive element selection, screenshot capture, and proper resource closure. There are no obvious gaps, as all core workflows from initiation to termination are addressed effectively.

    Maintenance

    ActivityInactive
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables visual annotation on web pages for Claude Code, allowing element selection, comment addition, screenshot capture, and structured UI feedback for code fixes via an MCP server.
      MIT
    • A
      license
      A
      quality
      C
      maintenance
      Enables visual browser feedback collection directly into Claude Code. Users can point at elements in their browser and send annotated feedback that Claude can act on immediately.
      12
      1
      MIT
    • A
      license
      Not graded
      quality
      A
      maintenance
      It gives Claude Code a real browser so it can navigate, interact with, screenshot, and test web applications from the terminal. It also supports assertions, session recording, visual regression checks, and network/console monitoring.
      106 npm
      MIT