proxlane

@proxlane/sdk

An MCP server that gives an AI agent a web-fetch tool that fails honestly.

A plain fetch hands an agent a 200 with a captcha in the body, and the agent summarises the captcha. This tool fetches through your own Proxlane gateway, which fails over between scraping APIs and checks every response for block pages. The agent gets the page as readable text, or a failure that says what happened and whether retrying is worth it. A block page is never returned as content.

Set up

You need a running gateway with at least one provider key. That is one container: proxlane.dev/docs/quickstart.

Then add the server to your MCP client. Claude Code:

claude mcp add proxlane --env PROXLANE_URL=http://localhost:8787 --env PROXLANE_API_KEY=your-gateway-key -- npx -y @proxlane/[email protected]

Claude Desktop, Cursor and other clients take the same three things in their JSON config:

{
  "mcpServers": {
    "proxlane": {
      "command": "npx",
      "args": ["-y", "@proxlane/[email protected]"],
      "env": {
        "PROXLANE_URL": "http://localhost:8787",
        "PROXLANE_API_KEY": "your-gateway-key"
      }
    }
  }
}

PROXLANE_API_KEY is the gateway's own key, not a provider's. Provider keys stay on the gateway. Pin the version, as above: npx -y otherwise runs whatever was published last, with that key in its environment.

Every call spends provider credits, several times more with render or wait_for. The tool is deliberately not marked read-only, so clients that auto-approve read-only tools still ask.

The tool

fetch_page

Argument
url	required	http or https
render	optional	run JavaScript first; costs several times more
wait_for	optional	a CSS selector to wait for; implies render, and the page is a failure if it never appears
country	optional	two-letter code, e.g. us
format	text (default) or html
max_chars	default 20000	longer results are truncated, and say so

On failure the result is marked as an error and reads:

{ "ok": false, "outcome": "SOFT_BLOCK", "class": "blocked", "retryable": true, "chain": "scraperapi:SOFT_BLOCK>brightdata:SOFT_BLOCK" }

class is one of six that never grows, and retryable follows from it: the site's own answer (target) and a bad request (client) are not worth retrying; a block or a provider problem may be, later. The gateway has already tried every provider it could, so an immediate retry is not.

In your own process

import { createServer } from '@proxlane/sdk';
import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';

await createServer({ gatewayUrl: 'http://localhost:8787', apiKey: process.env.PROXLANE_API_KEY! })
	.connect(new StdioServerTransport());

Apache-2.0.