mirror of
https://github.com/lukaszraczylo/mcp-captcharesolver.git
synced 2026-10-12 00:45:28 +00:00
73 lines
3.3 KiB
JavaScript
73 lines
3.3 KiB
JavaScript
// Reference: detect captchas on a page using Lightpanda (a lightweight headless browser)
|
|
// as the "hands" and captcha-solver-mcp's deterministic `detect_captcha` as the "brain".
|
|
//
|
|
// `detect_captcha` is a pure, deterministic HTML scan — it needs no LLM call to classify
|
|
// known providers. It flags reCAPTCHA v3 and Cloudflare Turnstile as `llm_solvable:false`
|
|
// because those are behavioral / proof-of-work challenges, not vision puzzles.
|
|
//
|
|
// This example shows the simplest possible integration: fetch the page DOM with Lightpanda,
|
|
// hand the HTML to the server, and print what was found. No clicking, no tokens.
|
|
//
|
|
// Run it yourself (this repo does not vendor node_modules):
|
|
// npm i playwright # used here purely as a CDP client to talk to Lightpanda
|
|
// lightpanda serve --host 127.0.0.1 --port 9222 # start Lightpanda's CDP endpoint
|
|
// node examples/lightpanda/detect.mjs https://your-own-test-page.example
|
|
//
|
|
// Required env:
|
|
// CAPTCHA_BIN=/abs/path/to/bin/captcha-solver-mcp
|
|
// CAPTCHA_LLM_PROVIDER=anthropic # any valid provider; detection won't call it
|
|
// CAPTCHA_LLM_API_KEY=sk-...
|
|
// CAPTCHA_LLM_MODEL=claude-opus-4-8
|
|
// LIGHTPANDA_CDP=ws://127.0.0.1:9222 # Lightpanda CDP websocket (optional override)
|
|
//
|
|
// AUTHORIZED USE ONLY. Detect captchas on pages you own or are permitted to test.
|
|
|
|
import { chromium } from 'playwright';
|
|
import { McpStdioClient } from '../mcp-client.mjs';
|
|
|
|
const TARGET_URL = process.argv[2];
|
|
const CAPTCHA_BIN = process.env.CAPTCHA_BIN || '../../bin/captcha-solver-mcp';
|
|
const LIGHTPANDA_CDP = process.env.LIGHTPANDA_CDP || 'ws://127.0.0.1:9222';
|
|
|
|
if (!TARGET_URL) {
|
|
console.error('usage: node detect.mjs <url-of-your-own-test-page>');
|
|
process.exit(2);
|
|
}
|
|
|
|
// 1. Start the brain. detect_captcha is deterministic, but the server still validates its
|
|
// LLM config at startup, so the CAPTCHA_LLM_* vars must be present and valid.
|
|
const mcp = new McpStdioClient(CAPTCHA_BIN, {
|
|
CAPTCHA_LLM_PROVIDER: process.env.CAPTCHA_LLM_PROVIDER,
|
|
CAPTCHA_LLM_BASE_URL: process.env.CAPTCHA_LLM_BASE_URL,
|
|
CAPTCHA_LLM_API_KEY: process.env.CAPTCHA_LLM_API_KEY,
|
|
CAPTCHA_LLM_MODEL: process.env.CAPTCHA_LLM_MODEL,
|
|
});
|
|
await mcp.initialize();
|
|
|
|
// 2. Connect to Lightpanda over CDP and fetch the rendered DOM.
|
|
// (Lightpanda speaks the Chrome DevTools Protocol, so any CDP client works; we reuse
|
|
// Playwright's connectOverCDP here to keep the example dependency-light.)
|
|
const browser = await chromium.connectOverCDP(LIGHTPANDA_CDP);
|
|
const context = browser.contexts()[0] || (await browser.newContext());
|
|
const page = context.pages()[0] || (await context.newPage());
|
|
|
|
await page.goto(TARGET_URL, { waitUntil: 'domcontentloaded' });
|
|
const html = await page.content(); // full serialized DOM, including captcha iframes/markers
|
|
|
|
// 3. Ask the brain to classify. Pure HTML scan — no LLM round-trip.
|
|
const result = await mcp.callTool('detect_captcha', { url: TARGET_URL, html });
|
|
|
|
console.log(result.summary);
|
|
for (const d of result.detected) {
|
|
const verdict = d.llm_solvable ? 'LLM-solvable' : 'DETECT-ONLY (behavioral/PoW)';
|
|
console.log(`- ${d.type}${d.sitekey ? ` sitekey=${d.sitekey}` : ''} -> ${verdict}`);
|
|
if (d.notes) console.log(` notes: ${d.notes}`);
|
|
}
|
|
|
|
if (result.detected.length === 0) {
|
|
console.log('no known captcha provider found in the DOM');
|
|
}
|
|
|
|
mcp.close();
|
|
await browser.close();
|