I built MCP Sub-50ms fetches for protected websites

#agents #mcp #showdev #webscraping

I built Karon.

A lot of "AI agent infrastructure" feels backwards to me: lots of abstraction, lots of demos, and not much that actually helps an agent reach a real page on the live web.

So I built something simpler: an API that fetches pages that normal HTTP clients often fail to reach, plus an agent-friendly endpoint that returns clean markdown/text instead of raw HTML.

The basic flow is:

The first request clears the WAF challenge (3–10s)
Session state gets cached
Subsequent requests to the same site reuse the session (~45ms)
/v1/agent/browse returns cleaned content for agent workflows

What I wanted was not a "platform" pitch, just infrastructure that is immediately usable.

Available now:

Free tier
No credit card
Failed requests are free
Google sign-in for self-serve API keys
MCP package for agent workflows: pip install karon-mcp
/v1/agent/browse for markdown / text extraction

Examples:

Raw HTML:

curl -X POST https://api.karonlabs.net/v1/fetch \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"url":"https://www.nike.com"}'

Clean markdown for agent workflows:

curl -X POST https://api.karonlabs.net/v1/agent/browse \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"url":"https://www.nike.com","extract":"markdown"}'

MCP (Claude / Cursor — no install needed, uvx handles it):

{
  "mcpServers": {
    "karon": {
      "command": "uvx",
      "args": ["karon-mcp"],
      "env": { "KARON_API_KEY": "YOUR_KEY" }
    }
  }
}

A few boundaries:

Ticketing, login automation, financial sites, and similar abuse-heavy categories are blocked by default
Rate limits are per key
Content is not stored beyond cache TTL

Links: