Screenshot API for AI Agents

Give your AI agent eyes. One MCP tool call returns a screenshot your agent can actually see.

→ Join the early access waitlist · Free tier included · Hosted API coming soon

What it does

AI agents can't see web pages — they get raw HTML, which is noisy, incomplete (no JS-rendered content), and full of boilerplate. This MCP server gives Claude, Cursor, and any MCP-compatible agent three native tools:

describe_page is the key differentiator: most agents need a screenshot and some text context. This saves a round-trip.

Related MCP server: Junipr MCP Server

Quick start (self-hosted, free)

Claude Desktop

// ~/Library/Application Support/Claude/claude_desktop_config.json
{
  "mcpServers": {
    "screenshot-api": {
      "command": "npx",
      "args": ["-y", "screenshot-api-mcp"]
    }
  }
}

Restart Claude Desktop. Then ask:

"Take a screenshot of https://news.ycombinator.com and summarise the top 5 stories"

Cursor

Settings → MCP → Add Server:

{
  "name": "screenshot-api",
  "command": "npx",
  "args": ["-y", "screenshot-api-mcp"]
}

Python (LangChain / CrewAI)

from langchain_mcp_adapters.client import MultiServerMCPClient

client = MultiServerMCPClient({
    "screenshot": {
        "command": "npx",
        "args": ["-y", "screenshot-api-mcp"],
        "transport": "stdio",
    }
})

tools = await client.get_tools()
# → [take_screenshot, describe_page, capture_pdf]

Why not just use the HTML?

• JS-rendered content — SPAs, lazy-loaded sections, and canvas elements are invisible to scrapers. Playwright renders everything.

• Stealth mode — ~40% of real URLs return bot-block pages to naive scrapers. Evasion is built in.

• Response cache — same URL twice in a loop? Second call is instant (5-min TTL, no double-billing).

• MCP-native — structured outputs designed for agent loops, not bolted-on HTTP wrappers.

Hosted API (coming soon)

The self-hosted version requires Node.js + a local Playwright/Chromium install (~300MB). The hosted API has none of that — your agent hits an endpoint, gets back the image.

Join the waitlist — early access gets 3 months at 40% off.

Planned tiers:

HTTP API (self-hosted)

GET /screenshot

curl "http://localhost:3099/screenshot?url=https://example.com" -o out.png

GET /health

{ "status": "ok", "browserConnected": true }

Local development

npm install
npm run dev        # tsx watch
npm test           # vitest + real Chromium (~60s)
npm run typecheck

Deploy to Fly.io (self-host at scale)

export PATH="$HOME/.fly/bin:$PATH"
flyctl auth login
flyctl launch --no-deploy --copy-config
flyctl secrets set PROXY_SECRET=$(openssl rand -hex 24)
flyctl deploy
flyctl checks list   # /health should be green

Memory: Fly's default 256MB will OOM Chromium. fly.toml is pre-set to 1GB. Don't scale below 512MB.

Add FLY_API_TOKEN to GitHub repo secrets to enable auto-deploy on push to main.

Configuration

All settings via environment variables. See .env.example.

Key vars: PORT (3099), MAX_CONCURRENCY (2), PROXY_SECRET (empty = auth off locally), NAV_TIMEOUT_MS (20000).

Roadmap

Phase A — Validation (now)

• MCP server with 3 tools

• Stealth / Cloudflare bypass

• Response cache (5-min TTL)

• Landing page + waitlist

• npm publish (screenshot-api-mcp)

• Smithery + mcp.so + Glama listings

• 20 waitlist signups ← gate to Phase B

Phase B — Hosted API (if gate passed)

• Lemon Squeezy billing (usage tiers)

• API key issuance + metering

• Fly.io deploy (1GB VM, min_machines_running=1)

• First paying customer

Phase C — Growth

• SEO content ("how to give Claude screenshots")

• Zapier integration

• LangChain / CrewAI tool package