How to Give Your AI Agent Real Web Access (Screenshot, Scrape, Extract JSON from Any URL)

#agents #ai #api #webscraping

Most AI agents hallucinate when you ask them about live web content. Not because they're broken — but because they don't actually have web access. They're reasoning from training data that may be months old.

Here's how we solved this at snapapi.pics: a single API that gives agents real-time access to any webpage — screenshot, scraped text, structured JSON extraction, PDF generation, and video recording.

The problem

When your Claude Code or Cursor agent needs to:

Check a competitor's pricing page
Extract structured data from a JS-rendered site
Take a screenshot of a deployment for visual QA
Generate a PDF of a web page

...it either hallucates, asks you to do it manually, or requires a whole Playwright/Puppeteer setup just to fetch a URL.

The solution: one API call

# Screenshot any URL
curl -X POST https://api.snapapi.pics/v1/screenshot \
  -H "X-Api-Key: YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"url": "https://example.com", "format": "png"}'

# Scrape page content as markdown
curl -X POST https://api.snapapi.pics/v1/scrape \
  -H "X-Api-Key: YOUR_KEY" \
  -d '{"url": "https://news.ycombinator.com"}'

# Extract structured JSON using a schema
curl -X POST https://api.snapapi.pics/v1/extract \
  -H "X-Api-Key: YOUR_KEY" \
  -d '{
    "url": "https://somesite.com/pricing",
    "schema": {
      "plan_name": "string",
      "price": "number",
      "features": "array"
    }
  }'

Using it in Claude Code via MCP

The real unlock for agents is our MCP server. Add it once to your Claude Code or Cursor config:

{
  "mcpServers": {
    "snapapi": {
      "command": "npx",
      "args": ["snapapi-mcp"],
      "env": { "SNAPAPI_KEY": "your_key_here" }
    }
  }
}

Now your agent can call screenshot, scrape, extract, pdf, and video as native tools — no manual fetching, no hallucinations.

Stack behind it

For those curious about the implementation: we run Node.js + TypeScript + Fastify v4, with Playwright (Chromium) for browser automation, BullMQ for job queuing, PostgreSQL + Redis for storage, and PM2 for process management. All browser jobs go async through the queue — critical for not timing out on heavy pages.

One hard-learned lesson: never process Playwright jobs synchronously in the request handler. Each browser context spikes 200-400MB. Queue them.

Free tier

200 requests/month free, no credit card. SDKs available in JavaScript, Python, Go, PHP, Swift, Kotlin.

→ snapapi.pics | npm: snapapi-mcp

Happy to answer questions about the architecture or the MCP integration in the comments.

DEV Community