io.github.mcpware/ui-annotator-mcp

Official

by mcpware · JavaScript

Annotates any web page with hover labels for AI assistants — zero extensions, any browser

Add hover labels to web pages for AI assistants without extensions or plugins.

Topics

accessibilityai-toolsclaudeclaude-codedeveloper-toolsfrontendmcpmcp-servermodel-context-protocolreverse-proxyweb-developmentanthropicautomationbrowser-automationdevtoolsllmllm-toolsopen-sourcetypescriptweb-scraping

README.md9 KBView on GitHub

UI Annotator MCP

English | 廣東話

Bridge the gap between what you see and what AI can reference — in any browser, zero extensions.

The only tool that puts visible labels on every web element. Hover any element, see its name. Tell your AI assistant "make the sidebar wider" — it knows exactly which element you mean. No screenshots, no CSS selectors, no miscommunication.

Demo

Why It Matters

Dramatically improves AI-driven UI design and iteration. The pain: telling AI "move that button next to the search bar" never works because the AI can't see your page. UI Annotator fixes this — hover over any element and its component name appears as a label. Now you say "move SearchButton below NavBar" and Claude edits the right component instantly. No browser extensions, works with any framework. The workflow becomes: open page → hover to identify elements → describe changes using real component names → Claude edits → refresh and repeat. Turns a frustrating back-and-forth into a fluid design loop.

The Problem

When reviewing a web UI with an AI coding assistant, the hardest part isn't the code change — it's describing which element you want changed.

"That thing on the left... the second row... no, the one with the icon..."

You don't know what it's called. The AI doesn't know what you're pointing at. You waste time on miscommunication instead of shipping.

The Solution

Open your page through the annotator proxy. Hover any element — instantly see its name, CSS selector, and dimensions. Now you both speak the same language.

code

# Start the MCP server
npx @mcpware/ui-annotator

# Open in ANY browser
http://localhost:7077/localhost:3847

That's it. No browser extensions. No code changes. No setup. Works in Chrome, Firefox, Safari, Edge — any browser.

How It Works

code

Your app (localhost:3847)
        │
        ▼
┌─────────────────────┐
│  UI Annotator Proxy  │ ← Reverse proxy on port 7077
│  (MCP Server)        │
└─────────────────────┘
        │
        ▼
Proxied page with hover annotations injected
        │
        ├──► User sees: hover overlay + tooltip with element names
        └──► AI sees: structured element data via MCP tools

The proxy fetches your page, injects a lightweight annotation script, and serves it back. The script scans the DOM, identifies named elements, and reports them to the MCP server. Your AI assistant queries the server to understand what's on the page.

Features

Hover Annotations

Hover any element to see:

Element name (pink) — the human-readable identifier
CSS selector (monospace) — the technical reference
Content preview — what text the element contains
Dimensions — width × height in pixels

Inspect Mode

Click the Inspect button in the toolbar (or let your AI toggle it). In inspect mode:

Click any element → copies its name to clipboard
All page interactions are paused (clicks don't trigger buttons/links)
Click Inspect again to return to normal mode

The toolbar sits at the top center of the page showing:

Inspect toggle button
Element count
Helpful subtitle explaining what to do
Collapse button (▲) to minimize when not needed

MCP Tools for AI

Tool	What it does
`annotate(url)`	Returns proxy URL for user to open in any browser
`get_elements()`	Returns all detected UI elements with names, selectors, positions
`highlight_element(name)`	Flash-highlights a specific element so user can confirm
`rescan_elements()`	Force DOM rescan after page changes
`inspect_mode(enabled)`	Toggle inspect mode remotely

Why Not Just Use DevTools?

	Browser DevTools	UI Annotator
Target user	Frontend developers who know the DOM	Anyone — QA, PM, designer, junior dev
Learning curve	Need to understand DOM tree, CSS selectors, box model	Hover and read — zero learning
Communication	"The `div.flex.gap-4` inside the second child of..."	"The `sidebar`"
Language	CSS/HTML technical terms	Human-readable names
Setup	Teach people to open DevTools + navigate the DOM	Open a URL
AI integration	None — AI can't see what you're inspecting	MCP — AI sees the same element names you do

DevTools is for debugging. UI Annotator is for communication — giving humans and AI a shared vocabulary for UI elements.

Why Not Use Existing Tools?

None of these do what UI Annotator does — live visual labels on every element via reverse proxy:

Tool	Approach	Why we're different
browser-use (82K⭐)	AI automation framework	Automates browsers, doesn't label elements for humans. Different use case entirely.
Chrome DevTools MCP (31K⭐)	DOM snapshot + element UIDs	AI can inspect, but humans don't see visual annotations. No shared vocabulary.
Playwright MCP (29K⭐)	Accessibility tree snapshot	Returns structured text, no visual overlay. Truncates important context.
OmniParser	Screenshot + CV detection	Screenshot-based, not live DOM. ~40% accuracy on hard benchmarks.
MCP Pointer (526 users)	Chrome extension + MCP	Requires Chrome extension. Human clicks to select — no hover overlay.
Agentation	npm embedded in your app	Requires code changes. React 18+ dependency. Not zero-config.
Vibe Annotations	Chrome extension	Extension-based, developer-only annotation workflow.

Feature Comparison

Feature	UI Annotator	MCP Pointer	Agentation	Cursor	Chrome DevTools MCP
Visual hover annotation	Yes	No	Partial	Yes (IDE only)	No
Shows element names	Yes	Yes	Yes	No (high-level)	Programmatic
Shows dimensions	Yes	Yes	Yes (Detailed)	Yes	Programmatic
MCP server	Yes	Yes	Yes	No	Yes
Zero browser extensions	Yes	No	Yes	N/A	No
Zero code changes	Yes	Yes	No	N/A	Yes
Any browser	Yes	Chrome only	Desktop only	Cursor only	Chrome only
Zero dependencies	Yes	Chrome ext	React 18+	Cursor	Chrome
Click to copy element name	Yes	No	No	No	No

Architecture

Zero external dependencies

Reverse proxy: Node.js built-in http module
MCP server: @modelcontextprotocol/sdk (stdio transport)
Communication: HTTP POST (browser → server) + GET polling (server → browser)
No WebSocket, no Express, no browser extension

How the proxy works

User requests localhost:7077/localhost:3847
Proxy fetches http://localhost:3847
For HTML responses:
- Injects fetch() / XMLHttpRequest interceptor (rewrites API paths through proxy)
- Rewrites href="/..." and src="/..." attributes to route through proxy
- Injects annotation script before </body>
For non-HTML (CSS, JS, images): passes through directly
Strips Content-Security-Policy headers to allow injected script

How annotation works

Script scans DOM for elements with id, class, semantic roles, or interactive roles
On hover: positions overlay border (follows border-radius) + positions tooltip (always within viewport)
Reports all detected elements to server via POST /__annotator/elements
Polls GET /__annotator/commands every second for server instructions (highlight, rescan, inspect toggle)
MutationObserver auto-rescans when DOM changes

Quick Start

With Claude Code

bash

# Add as MCP server
claude mcp add ui-annotator -- npx @mcpware/ui-annotator

# Then in conversation:
# "Annotate my app at localhost:3847"
# → AI returns proxy URL, you open it, hover elements, discuss changes by name

Manual