웹 검색 도구 만 교육 사용.

왜 크롬 크롤러 mcp?

인기있는 Playwright-mcp를 사용 했습니까? 그것은 ** 자신의 프로필을 재사용하고 상태에 로그인 **.

• 여기서 mcp는 브라우저 인스턴스를 생성하지 않고 스스로 행동합니다.

웹 에이전트 mcp return huge 페이지 컨텐츠를 llm에 경험해 보시고, 컨텍스트 창에서 빠른 실행을 합니다.

• 여기에 우리는 원래 페이지와 상호 작용 할 수있는 능력을 잃지 않고 페이지 콘텐츠를 집계.

llm에 대해 걱정할 필요가 있습니다. 수백 개의 웹 페이지 처리 ** 아무것도 같은 토큰**.

• LLM을 사용하지 않고도 crawling을 실행하기 위해, 똑똑똑하게 speical parsers 및 워크플로우를 만들 수 있습니다. 당신의 대리인은 본래 조사 및 쓰기 parser/workflow를 위해만 필요합니다.

확장을 사용 하 여 걱정 ** 당신의 개인 정보 보호 ** AI 회사의 브라우저 도구를 사용 하 여.

• 여기에 당신은 훨씬 빠르고 강력한 방법으로 크롤러를 소유하고 있습니다 (예, vibe 코딩은 당신이 필요로하는 것입니다), 그냥 귀하의 개인 정보를 포기하지 않고 그것을 실행

Context 폭발 문제

MCP를 긁는 전통적인 웹은 중요한 도전을 직면 : ** 빨리 LLM 컨텍스트 윈도우를 배출하는 전체 HTML 콘텐츠를 반환 **.

Chrome-Crawler-mcp

Educational Use Only web browsing tools.

Why Chrome-Crawler-mcp?

Have you used popular playwright-mcp? it is hard to reuse your own profile and log in status.

• Here, mcp act as yourself without creating seperate browser instance.

Have you experience the web agent mcp return huge page content to llm and you agent quicky run out of context window.

• Here we condense page content without losing the ability to interacte with original page.

Have you worry about asking llm to process hundreds of webpages which burn your tokens like nothing.

• Here you can ask your agent to smartly create speical parsers and workflow to run crawling without using LLM. Your agent is only needed for original investigation and writing parser/workflow for you.

Do you hate to use extensions and worry about give up your privacy of using AI companies’ browser tools.

• Here you build you own crawler in much faster and robust way (yes, vibe coding is what you need), and just run it without giving up your privacy

The Context Explosion Problem

Traditional web scraping MCPs face a critical challenge: they return full HTML content that quickly exhausts LLM context windows. A typical modern webpage can contain 50,000+ tokens of HTML, leaving little room for reasoning, tool calls, or multi-turn conversations. This forces agents to either:

• Work with severely limited context
• Repeatedly re-fetch the same content
• Miss important content due to truncation

Chrome-Crawler-mcp solves this with intelligent HTML sanitization that reduces pages to ~5-10% of their original size while preserving all interactive capabilities.

Our Solution: Smart Sanitization + Element Remapping

1. Token-Efficient HTML Sanitization (src/html_sanitizer.py:14)

• Removes scripts, styles, SVGs, and non-essential elements
• Strips redundant attributes and normalizes whitespace
• Compresses typical 50KB pages to 5-8KB (~10x reduction)
• Result: Agents can now work with 10+ pages in a single context window

2. Intelligent Element ID Assignment

• Despite heavy sanitization, agents still need to click buttons and fill forms
• Solution: We assign unique data-web-agent-id attributes to every interactable element before sanitization
• These IDs are preserved through the sanitization process and injected back into the actual DOM
• Agents interact with elements using stable IDs, not fragile XPath or CSS selectors
• Result: Robust element targeting that survives page updates and dynamic content

3. Multi-Strategy Locator Fallbacks

• Each element gets XPath, CSS selector, ID, class, and href-based locators
• If the primary web-agent-id lookup fails, we automatically fall back to alternative strategies
• Result: 99%+ success rate for element interaction

Delegation Over Reasoning: Special Parsers

The Problem with LLM-Based Parsing: Most web agents ask the LLM to:

1. Read raw HTML (20,000+ tokens)