For AI agents · Model Context Protocol
SnapshotFlow MCP Server
Give your AI agent a screenshot tool it can call directly. The SnapshotFlow MCP server exposes SnapshotFlow's rendering engine over the Model Context Protocol, so coding agents and AI assistants can capture, batch, and compare web pages and read them as Markdown without scraping HTML or scripting a browser.
Example conversation with the SnapshotFlow MCP server
Take a screenshot of https://stripe.com and tell me what a visitor sees on the first screen.
Request
{
"url": "https://stripe.com",
"width": 1280,
"height": 800
}

The first screen leads with "Financial infrastructure to grow your revenue" and two calls to action: Get started and Sign up with Google. Below them is a strip of customer logos.
One thing to note: a "11 sales reps available" chat widget covers the right end of the logo strip on a 1280×800 screen.
Capture the home pages of github.com, stripe.com, vercel.com and en.wikipedia.org and compare their main call to action.
Request
{
"urls": [
"https://github.com",
"https://stripe.com",
"https://vercel.com",
"https://en.wikipedia.org"
]
}




- GitHub: an email field joined to Sign up for GitHub, plus a secondary Download GitHub Copilot app.
- Stripe: Get started next to Sign up with Google.
- Vercel: Deploy now with Talk to sales as the alternative.
- Wikipedia: no sign-up push; search and reading come first.
Read the Wikipedia main page as Markdown and tell me what is featured today.
Request
{
"url": "https://en.wikipedia.org/wiki/Main_Page",
"format": "markdown"
}
# Welcome to Wikipedia,
the free encyclopedia that anyone can edit.
260,720 active editors · 7,240,037 articles in English
## From today's featured article
**Menches** was the village scribe of Kerkeosiris
in Ptolemaic Egypt. …Today's featured article is about Menches, a village scribe in Ptolemaic Egypt whose papyri, found in 1900, are the only known evidence of his life. The page also lists 7,240,037 English articles.
No image was needed, so this cost far fewer model tokens than a screenshot.
Screenshots are real SnapshotFlow captures; the requests and replies are abbreviated examples. More ready-made tasks are in the prompt library.
Three ways to connect an agent
Hosted MCP server
Add one URL to your AI client and sign in with OAuth. Nothing to install, no API key in the config.
Pick a client →Agent Skill
A SKILL.md that teaches an agent to call the REST API with your API key. Save it to your agent's skills folder, for example .claude/skills/snapshotflow/.
API and SDKs
Call the API from your own agent product or pipeline when a screenshot is a fixed step, not a decision the model makes.
Read the API docs →Connect your agent
The remote MCP endpoint is served directly by the SnapshotFlow backend over Streamable HTTP:
https://api.snapshotflow.com/mcp
Most clients that use a JSON config take one entry. Some, such as Claude Code and VS Code, also need "type": "http"; the client guides below show the exact format.
{
"mcpServers": {
"snapshotflow": {
"url": "https://api.snapshotflow.com/mcp"
}
}
}
The endpoint is protected by OAuth 2.0. On first connect, your MCP client opens a browser to sign in and is issued a short-lived Bearer token automatically — no API key header to manage. Requests without a valid token return 401.
Tool calls run on SnapshotFlow servers, so there is nothing to install locally. It also means pages must be publicly reachable: SnapshotFlow cannot open localhost on your machine. Use a preview deployment, or pass the page's HTML with the html parameter.
Choose your client
claude mcp add command, then sign in from /mcp.
Setup guide →
Cursor
Add it to .cursor/mcp.json and use it from Agent.
Setup guide →
Claude
A custom connector for Claude on the web, desktop, and mobile.
Setup guide →
Codex
codex mcp add and codex mcp login for the CLI, IDE, and app.
Setup guide →
VS Code & GitHub Copilot
.vscode/mcp.json plus Copilot agent mode.
Setup guide →
serverUrl in mcp_config.json.
Setup guide →
Zed
A remote context server with OAuth from the agent panel.
Setup guide →
gemini mcp add --transport http snapshotflow https://api.snapshotflow.com/mcp, then /mcp auth snapshotflow.
Available MCP tools
Every tool and its full input schema is described in the MCP server card. Your AI client spends its own model tokens, billed by your AI provider; SnapshotFlow only counts screenshots.
| Tool | What it does | Counts as |
|---|---|---|
screenshot | Capture a URL or HTML as PNG, JPEG, or WebP, with optional page metadata and Markdown content. | 1 screenshot; cache hits free |
batch_screenshot | Capture 1–10 URLs in one parallel request, with ads and cookie banners blocked by default. Use include_images: false for links only, or async: true to run it as a job. | 1 per captured URL; cache hits free |
visual_diff | Pixel-compare two URLs and return a highlighted diff image with change statistics. | 2 screenshots |
extract_text | Return a page as Markdown, plain text, or HTML. Far fewer LLM tokens than an image when you only need the words. | 1 screenshot |
check_job | Check an async job started from MCP or the HTTP API, including batch progress and partial results. | Free |
New accounts get 300 free screenshots for the lifetime of the account, no credit card required. See pricing for paid plans.
Control the page before your agent sees it
Real pages load late, show overlays, and change with the viewport. The screenshot tool takes options that give the agent a clean, predictable image. Agents pick them from the tool schema, or you can name them in the prompt.
Wait for the right state
wait_for_selector waits for an element to appear; delay adds a fixed pause for animations.
Remove visual noise
block_ads and block_cookie_banners keep ads and consent pop-ups out of the image.
Capture the right area
width and height set the viewport, full_page captures the whole scroll height, and dark_mode renders the dark theme.
Get text with the image
extract_content returns the page as Markdown and metadata returns title and Open Graph data in the same call.
What agents use it for
A screenshot shows how a page looks; extracted text shows what it says. With both, an agent can check, compare, and report on live pages instead of guessing from raw HTML.
Machine-readable discovery
The MCP server card lets agents and registries discover SnapshotFlow's tools automatically:
https://snapshotflow.com/.well-known/mcp/server-card.json
It follows the MCP server-card schema and lists each tool's name, description, and JSON input schema.
Make your AI agent web-aware
Connect the SnapshotFlow MCP server in minutes. Free tier: 300 screenshots for the lifetime of the account. No credit card.
Frequently asked questions
What is the SnapshotFlow MCP server?
It is a Model Context Protocol server that exposes SnapshotFlow's screenshot engine as native tools — screenshot, batch_screenshot, visual_diff, extract_text, and check_job — so AI agents like Claude and Cursor can capture, compare, and read web pages directly.
Which MCP endpoint should I use?
The remote MCP endpoint is https://api.snapshotflow.com/mcp. Tool calls run on SnapshotFlow servers; nothing renders on your machine.
What does an MCP tool call cost?
Your AI client's model tokens are billed by your AI provider, not by SnapshotFlow. SnapshotFlow counts screenshots: screenshot and extract_text count as 1, visual_diff as 2, batch_screenshot as 1 per captured URL, and check_job is free. Repeat captures served from cache are free.
Is the SnapshotFlow MCP server free?
New accounts include 300 screenshots for the lifetime of the account. No credit card is required.
Do I need an API key to use MCP?
No. The MCP server uses OAuth: your client opens a browser sign-in and receives a short-lived token. API keys are only needed for the REST API and SDKs.
Does SnapshotFlow MCP work with Claude Desktop and Claude Code?
Yes. In Claude on the web, desktop, and mobile, add SnapshotFlow as a custom connector (Claude guide). In Claude Code, run claude mcp add --transport http snapshotflow https://api.snapshotflow.com/mcp and sign in from /mcp (Claude Code guide).
MCP or the API: which should I use?
Use MCP when you work through an AI client such as Claude Code, Cursor, or Claude and want the model to decide when to capture a page. Use the API or SDKs when screenshots are a fixed step in your own code and you need full control over requests and responses.
Why give an agent a screenshot instead of the page text?
Text shows what a page says; a screenshot shows what visitors see, including layout, overlays, missing images, and broken styles. Use extract_text when only the words matter, since it costs far fewer LLM tokens, and a screenshot when the visual result matters.
Can SnapshotFlow capture localhost?
Not directly. Pages are rendered on SnapshotFlow servers, which cannot reach your machine. Use a preview deployment or tunnel URL, or pass the page's HTML with the screenshot tool's html parameter.
Is there a machine-readable MCP server card?
Yes. The MCP server card is published at /.well-known/mcp/server-card.json and describes every tool and its JSON input schema.
What OAuth permissions are required?
SnapshotFlow requests a single mcp scope, which lets the client call the MCP tools on your account. Review the sign-in prompt before you approve it.
Learn more
- What is the best Screenshot API for AI agents? — how MCP and WebMCP compare
- WebMCP explained — in-browser tools vs the remote MCP server
- Playwright MCP screenshots, Chrome DevTools MCP screenshots, and agent-browser screenshots — when a local browser is the better fit
- MCP tool deep-dive: extracting text 10× cheaper than a screenshot
- MCP prompt library — copy-ready prompts grouped by task
- Changelog: MCP setup guides for seven AI clients
- Changelog: async batch jobs and new batch_screenshot options
- MCP server reference — endpoints, auth, and the full tool list