claude-vision-mcp
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@claude-vision-mcp Screenshot the current webpage at localhost:3000"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Claude Vision MCP
A Model Context Protocol server that gives Claude Code the ability to see — capturing the screen, individual windows, screen regions, and live web pages, plus comparing screenshots for visual-regression checks.
Built so an AI coding agent can verify its own visual output: render a UI change, screenshot it, and look at the result instead of guessing.
Features
Six tools, exposed over MCP via FastMCP:
capture_screen— screenshot the full screen or a specific display, with token-aware downscaling.capture_window_tool— screenshot a single window by (partial, case-insensitive) title or app name.capture_screen_region— screenshot an arbitrary rectangular region by coordinates.capture_webpage_tool— screenshot any URL (includinglocalhostdev servers) via a headless Playwright browser; supports full-page capture and waiting on a CSS selector.compare_screenshots— diff two images and report the percentage and region of changed pixels (visual regression).list_windows— list all open windows asApplication | Window Titleto find a capture target.
Related MCP server: Browsershot MCP
Requirements
Python ≥ 3.10
macOS (window capture uses AppleScript; screen capture needs Screen Recording permission)
Dependencies:
mcp[cli],Pillow,playwright,numpy
Install
git clone https://github.com/wonderstone843/claude-vision-mcp.git
cd claude-vision-mcp
pip install -e .
playwright install chromium # only needed for capture_webpage_toolGrant Screen Recording permission to your terminal in System Settings → Privacy & Security → Screen Recording.
Use with Claude Code
Register the server (stdio):
claude mcp add claude-vision -- claude-vision-mcpOr add it to your MCP config manually:
{
"mcpServers": {
"claude-vision": {
"command": "claude-vision-mcp"
}
}
}Then ask Claude to, e.g., "screenshot localhost:3000 and check the hero section renders," or "capture the Blender window."
Project layout
claude_vision_mcp/
server.py # FastMCP server + the 6 tool definitions
capture.py # full-screen / window / region capture
windows.py # AppleScript window enumeration
browser.py # Playwright headless webpage capture
compare.py # pixel-diff comparisonLicense
MIT — see LICENSE.
Author: Joshua Penn
Available Tools
6 toolscapture_screenA
Capture a screenshot of the full screen or a specific display.
Parameters:
display: Display number (0 = all displays combined, 1 = main, 2 = secondary). Default 0.
max_size: Maximum pixel dimension for the largest side. Default 800. Smaller = faster + fewer tokens. Use 400 for quick checks, 1200 for detail.
Returns the screenshot as an Image. Requires Screen Recording permission in macOS System Settings.
Inputs: display, max_size.
capture_screen_regionA
Capture a specific rectangular region of the screen.
Parameters:
x: Left edge coordinate in pixels (0 = left edge of screen).
y: Top edge coordinate in pixels (0 = top edge of screen).
width: Width of the region in pixels.
height: Height of the region in pixels.
max_size: Maximum pixel dimension for the output. Default 800.
Use capture_screen first to see the full screen and identify coordinates. Returns the captured region as an Image.
Inputs: x (required), y (required), width (required), height (required), +1 more.
capture_webpage_toolA
Capture a screenshot of a web page using a headless browser.
Useful for capturing local dev servers (localhost), web apps, or any URL. The browser launches on first use and stays running for subsequent captures.
Parameters:
url: The URL to capture (e.g., "http://localhost:3000", "https://example.com").
viewport_width: Browser viewport width in pixels. Default 1280.
viewport_height: Browser viewport height in pixels. Default 720.
max_size: Maximum pixel dimension for the output. Default 800.
full_page: If true, capture the entire scrollable page. Default false.
wait_for_selector: Optional CSS selector to wait for before capturing (e.g., "#main-content", ".loaded"). Leave empty to skip.
Returns the web page screenshot as an Image.
Inputs: url (required), max_size, full_page, viewport_width, +2 more.
capture_window_toolA
Capture a screenshot of a specific window by its title.
Searches for a window whose title contains the given text (case-insensitive). Also matches against application name. Use list_windows() first to see available windows.
Parameters:
window_title: Full or partial window title to match (e.g., "Firefox", "Terminal", "Blender"). Also matches application names.
max_size: Maximum pixel dimension for the largest side. Default 800.
Returns the screenshot as an Image. If no match is found, returns an error listing available windows.
Inputs: window_title (required), max_size.
compare_screenshotsA
Compare two screenshots and describe visual differences.
Analyzes pixel differences between two images and reports the percentage of changed pixels and the region of change.
Useful for:
Verifying UI changes after code edits
Checking before/after visual regression
Confirming rendering output differences
Parameters:
image_path_1: Absolute path to the first (baseline) image.
image_path_2: Absolute path to the second (comparison) image.
max_size: Maximum pixel dimension for analysis. Default 800.
Returns a text description of the differences found.
Inputs: image_path_1 (required), image_path_2 (required), max_size. Returns: result
list_windowsA
List all open windows with their application names and titles.
Use this to find the correct window title before calling capture_window_tool. Returns a formatted list showing 'Application | Window Title' for each open window.
No parameters required.
Inputs: no named parameters. Returns: result
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 6 tools
Each tool has a clearly distinct purpose: capturing different screen sources (full, region, webpage, window), comparing images, or listing windows. No overlapping functionality.
Most tools follow a 'capture_' prefix pattern, but capture_webpage_tool and capture_window_tool have an inconsistent '_tool' suffix, and compare_screenshots and list_windows use different verb forms.
6 tools is well-scoped for a vision capture server, covering core capture methods plus comparison and window listing without excess.
Covers all major capture scenarios (screen, region, window, webpage) and adds comparison and window listing. Minor gap: no tool to list displays or get screen dimensions, but all essential workflows are present.
Related MCP Connectors
Capture screenshots, detect visual regressions between page versions, and analyze with AI.
Let your AI agent see any public web page: one screenshot tool, returns the page as an image.
Give your agent eyes: screenshots of any web page it can see, pages as markdown, and PDFs.
Screenshot, PDF and HTML-to-image rendering API so Claude and Cursor can see any web page.
Related MCP Servers
- AlicenseDqualityDmaintenanceEnables AI assistants to capture and analyze screenshots using Claude Vision API, providing AI-powered insights about desktop interface content, UI elements, and visual layouts.2101 npm2MIT
- AlicenseAqualityDmaintenanceEnables Claude Code to capture and analyze web page screenshots, responsive layouts, and page metadata using Puppeteer. It allows developers to perform visual UI inspections and compare designs across various viewports directly within the terminal.3MIT
- AlicenseNot gradedqualityDmaintenanceEnables screenshot capture and visual analysis using cloud or local vision models, with tools to describe screens, list windows, and analyze images.13MIT
- FlicenseNot gradedqualityAmaintenanceEnables AI vision workflows to capture fresh webpage screenshots, specific DOM elements, and timed screencasts, automatically limiting resolution and tiling full pages into 1072x1072 chunks for LLM-friendly processing.-