Use GetScreenshot in Claude Code
Add the GetScreenshot MCP server with one claude mcp add command. Claude Code can then capture any public page, render HTML it just wrote, and compare before and after.
Works with any GetScreenshot API key on any plan, from $5/month.
What you get
Five tools that let the agent look at web pages
Look at any public page
The agent passes a URL to capture_screenshot. A hosted browser renders the page and the tool returns images sized for vision models, plus a link to the full-resolution file.
Render HTML it just wrote
render_html takes an HTML string and returns a screenshot of it. The agent can check a component or email template without a local browser.
Compare before and after
compare_screenshots diffs two pages or two earlier captures and reports the share of changed pixels, where the change is, and a diff image.
Limits worth knowing: captures run on our servers, so they cannot reach localhost or private networks. Each call is a single capture, not a logged-in browsing session. For multi-step browser automation see GetScreenshot vs Playwright MCP. Screenshot links are public URLs that stay available for 30 days.
Before You Start
Three things to have ready
A GetScreenshot API key
Any plan works. Plans start at $5/month for 2,500 screenshots. Get a key.
Claude Code
Claude Code installed and signed in.
Node.js 20.9 or newer
Only for the local option, which runs npx -y getscreenshot-mcp on your machine. The hosted option needs nothing installed.
How to Set Up
Hosted server at https://mcp.getscreenshotapi.com/mcp, or a local server started with npx
Claude Code supports both connection styles. Use one of them, not both. Both give the agent the same five tools and use the same API key and quota.
Get your API key
Pick a plan on the pricing page. The key arrives by email after checkout. The MCP server works with any API key on any plan, including Lincoln at $5/month for 2,500 screenshots. Code FREE5 makes the first month of the $5 plan free. If you lose the key, recover it from your account page.
View pricing plansOption A: connect to the hosted server
Nothing to install. Run this in a terminal, with your key in place of YOUR_API_KEY.
claude mcp add --transport http getscreenshot https://mcp.getscreenshotapi.com/mcp \
--header "Authorization: Bearer YOUR_API_KEY"By default Claude Code saves the server in local scope: it is stored in ~/.claude.json, loads only in the project you ran the command from, and is not shared with anyone. Add --scope user to load it in all your projects.
Option B: run the server locally
Claude Code starts the server on your machine with npx. Needs Node.js 20.9 or newer.
claude mcp add --env GETSCREENSHOT_API_KEY=YOUR_API_KEY --transport stdio getscreenshot \
-- npx -y getscreenshot-mcpClaude Code's own options go before the server name. The -- separates them from the command that starts the server. The same scope rules apply as in option A.
Optional: share the server with your team, not the key
Project scope stores the server in a .mcp.json file at the project root, which is meant to be committed. Do not put the key in that file. Claude Code expands ${VAR} references in command, args, env, url and headers, so commit this instead:
.mcp.json
{
"mcpServers": {
"getscreenshot": {
"type": "http",
"url": "https://mcp.getscreenshotapi.com/mcp",
"headers": {
"Authorization": "Bearer ${GETSCREENSHOT_API_KEY}"
}
}
}
}Each person sets GETSCREENSHOT_API_KEY in their own shell. Claude Code asks each person to approve a project-scoped server before it uses it.
Do not run claude mcp add --scope project with a literal key in --header or --env. That writes the key into .mcp.json.
Confirm it is connected
List your servers from a terminal. getscreenshot should show Connected.
claude mcp listInside a Claude Code session, run /mcp to see the server status. claude mcp get getscreenshot prints the details of the saved entry.
Try a prompt
Ask for a screenshot of a public page. Approve the tool call when Claude Code asks for permission.
Take a screenshot of https://example.com and describe what is above the fold.Prompts to Try
Replace the example URLs and file names with your own
Check a deployed change
Take a full-page screenshot of https://example.com/pricing and list anything that looks broken or misaligned.
Check HTML you just wrote
Render the HTML in src/emails/welcome.html with render_html and tell me whether the header and button look right.
Compare staging and production
Compare https://staging.example.com with https://www.example.com and describe what changed.
Check a mobile layout
Capture https://example.com with the iphone12 device and tell me if the navigation fits on screen.
Each page captured uses one screenshot from your monthly quota. A comparison uses one per page it has to capture.
Tools the Server Adds
All five are read-only. Full inputs and outputs are in the MCP server reference
| Tool | What it does |
|---|---|
| capture_screenshot | Capture a public web page as images the model can see. Options cover full-page capture, a single element by CSS selector, device presets and wait conditions. |
| render_html | Render an HTML string (up to 2 MB) and screenshot it. Useful for a component or page the agent just wrote. |
| compare_screenshots | Visual diff of two pages or two earlier screenshots. Returns the changed-pixel percentage, the bounding box of the change and a diff image. |
| capture_pdf | Save a page as an A4 PDF and return a link that works for 30 days. |
| get_usage | Screenshots used, monthly limit and reset date for your key. Costs 1 quota unit. |
Troubleshooting
Common issues and how to resolve them
Server shows "Failed to connect", or tools say no API key is configured
- • Check the key for typos or extra spaces. Remove the entry with
claude mcp remove getscreenshotand add it again. - • Hosted setup: the header must read
Authorization: Bearer YOUR_API_KEY, with the wordBearerand a space before the key. - • With a shared
.mcp.json:claude mcp listprints a missing-variable warning whenGETSCREENSHOT_API_KEYis not set in your shell. - • If you edit JSON by hand, a remote entry needs
"type": "http". Claude Code reads an entry with aurland notypeas a configuration error and skips it.
Server shows "Pending approval"
- • That is a project-scoped server from
.mcp.jsonyou have not approved yet. Runclaudein the project and approve it.
A tool call times out on a slow page
- • Ask the agent to retry with
wait_until: "domcontentloaded". It is the fastest setting and the most reliable on heavy sites. - • A capture usually takes around 9 seconds. The API gives up after 90 seconds.
- • Failed captures still count against your monthly quota, so change the settings before retrying more than once.
The agent cannot capture localhost or an internal site
- • The capture runs on our servers, which cannot reach
localhostor private networks. - • For markup the agent just wrote, use
render_htmlwith the HTML itself. Inline the CSS and images or set<base href>, because relative asset URLs will not resolve. - • For a running app, give the agent a public preview URL or a tunnel address.
The image is too large for the conversation
- • Ask for
return: "url_only"to get thescreenshot_urlwithout inline images. - • Capture one part of the page with
elementand a CSS selector instead of the full page. - • Lower
max_slices(default 4) to limit how many images a full-page capture returns. - • Claude Code limits MCP tool output to 25,000 tokens by default. Set the
MAX_MCP_OUTPUT_TOKENSenvironment variable before starting Claude Code if you need more.
Still stuck? Email support@rasterwise.com.
Other AI Clients
The same server works in other MCP clients. See all options on the AI Agents page
Prefer instructions the agent loads on demand? See the GetScreenshot agent skill.
GetScreenshot is an independent product. It is not affiliated with or endorsed by the makers of Claude Code. Product names belong to their owners.
Get an API Key and Connect Claude Code
Plans start at $5/month for 2,500 screenshots. The MCP server works with any API key on any plan.