crawl4ai-mcp
A minimal MCP server for agent-friendly web extraction and search. Offers two tools: fetching real pages with Playwright and Crawl4AI, and searching across 7 engines with automatic fallback.
README
crawl4ai-mcp
<div align="center">
A minimal MCP server for agent-friendly web extraction and search.
Two tools: fetch real pages with Playwright + Crawl4AI, or search across 7 engines with automatic fallback.
</div>
Quick entry
| Audience | Read this |
|---|---|
| Human developer | README.zh-CN.md / README.md |
| Living in the AI era, delegating your remaining sanity to an agent | README_AGENT.md |
At a glance
| Item | Reality in this repo |
|---|---|
| MCP tools | 2 tools: fetch_urls + search_web |
| Single-page fetch | urls: ["https://example.com"] |
| Web search | search_web(query="...", engine="auto") — 7 engines, auto fallback |
| Search engines | DuckDuckGo · Bing · Google · Yandex · Sogou · 360Search · Baidu |
| Output | title + content + links + blocked + llm_used/llm_error |
| Non-LLM mode | First-class, default, usable without any model |
| LLM mode | Off by default. Enabled only with use_llm=true + optional llm_instruction |
| Fallback | Missing/failed LLM call automatically falls back to non-LLM result |
| Anti-bot realism | proxy / cookies / persistent profile / randomized browser behavior |
| License | AGPL-3.0-or-later |
How it works
Fetch flow:
---
config:
theme: base
themeVariables:
primaryColor: "#FAF9F5"
primaryTextColor: "#1A1A1A"
primaryBorderColor: "#D97757"
lineColor: "#8B5E3C"
secondaryColor: "#F5F1E8"
tertiaryColor: "#FAF9F5"
fontFamily: "ui-sans-serif, system-ui, -apple-system, sans-serif"
fontSize: "14px"
---
flowchart LR
A[URL list] --> B[Playwright + Crawl4AI]
B --> C{Fast path enough?}
C -- Yes --> D[Markdown / HTML]
C -- No --> E[Stronger fallback]
E --> D
D --> F{use_llm?}
F -- No --> G[Return result]
F -- Yes --> H[OpenAI-compatible cleanup]
H --> I{LLM success?}
I -- Yes --> J[Return enhanced result]
I -- No --> G
Search flow:
---
config:
theme: base
themeVariables:
primaryColor: "#FAF9F5"
primaryTextColor: "#1A1A1A"
primaryBorderColor: "#D97757"
lineColor: "#8B5E3C"
secondaryColor: "#F5F1E8"
tertiaryColor: "#FAF9F5"
fontFamily: "ui-sans-serif, system-ui, -apple-system, sans-serif"
fontSize: "14px"
---
flowchart LR
A[query + engine] --> B{engine=auto?}
B -- Yes --> C[Detect language]
C --> D[Build engine plan]
B -- No --> E[Use specified engine]
D --> F[Try engines in order]
E --> F
F --> G{Results?}
G -- Yes --> H[Aggregate + deduplicate]
G -- No, next engine --> F
H --> I[Return results]
Installation
Quick install (recommended)
Step 1: Create a virtual environment
# macOS/Linux - using system Python 3 (3.10-3.13)
python3 -m venv crawl4ai
source crawl4ai/bin/activate
# Windows
python -m venv crawl4ai
crawl4ai\Scripts\activate
Step 2: Install
pip install --upgrade pip
pip install crawl4agent
playwright install chromium
Alternative methods
If python3 is too old (3.9 or below):
# Use specific Python version (3.10, 3.11, 3.12, or 3.13)
python3.12 -m venv crawl4ai
source crawl4ai/bin/activate
pip install crawl4agent
Using conda:
conda create -n crawl4ai python=3.12
conda activate crawl4ai
pip install crawl4agent
playwright install chromium
Using pipx (global command):
pipx install crawl4agent
crawl4ai-mcp --help
Troubleshooting
Problem: "pip install" uses Python 2.7
# macOS: use python3 explicitly
python3 -m pip install crawl4agent
# Or check which pip you're using
which pip
pip --version
Problem: "No matching distribution found for crawl4agent"
- Check Python version:
python3 --version(must be 3.10-3.13) - Upgrade pip:
python3 -m pip install --upgrade pip
Problem: "playwright install" fails
- Use mirror (China):
export PLAYWRIGHT_DOWNLOAD_HOST=https://npmmirror.com/mirrors/playwright/ - Then:
python3 -m playwright install chromium
Why this project exists
Most generic “web fetch” tools either fail on JS-heavy pages or return too much boilerplate. This project focuses on four things:
- Non-LLM quality first: usable even with zero model config
- Minimal MCP surface: easier for agents, easier to maintain
- Pragmatic anti-bot workflow: proxy / cookies / persistent profile are first-class
- Golden regression review: full markdown outputs can be saved and inspected page by page
Core capabilities
Non-LLM mode
| Capability | Actual behavior |
|---|---|
| Rendering | Real browser rendering via Playwright |
| Extraction | Crawl4AI markdown/html extraction |
| Fallback | Fast path → stronger path when content is too thin |
| Cleanup | Remove obvious noise, compress blanks, strip data-image placeholders |
| Site tuning | Medium / Claude Docs / GitHub and other mainstream sites |
| ChatGPT shared links | Full conversation extraction from chatgpt.com/share/... URLs |
| Video transcripts | YouTube / Bilibili URLs prefer subtitle extraction via yt-dlp, then fall back to webpage extraction |
| Block detection | blocked=true for likely verification/interstitial output |
| Batch control | Bounded concurrency via concurrency |
Optional LLM mode
| Input | Meaning |
|---|---|
use_llm=true |
Turn on post-cleanup with an OpenAI-compatible model |
llm_instruction |
Tell the model what to keep / remove |
Important reality check:
- With
llm_instruction, the prompt is constraint-heavy and biased toward preserving original lines. - Without
llm_instruction, the model does a more generic “clean readable markdown” pass. - If the LLM call fails for any reason, the tool returns the original non-LLM extraction plus
llm_used=falseandllm_error.
MCP Tools
fetch_urls
{
"urls": ["https://a.com", "https://b.com"],
"format": "markdown",
"max_chars": 200000,
"concurrency": 3,
"use_llm": false,
"llm_instruction": "keep only the tutorial body and in-body references"
}
Use a single-element list if you only need one page.
For supported video URLs (youtube.com, youtu.be, bilibili.com, b23.tv), fetch_urls prefers transcript extraction and returns readable markdown built from subtitles when available.
Return shape
| Field | Meaning |
|---|---|
url |
Original URL |
final_url |
Final resolved URL after redirects |
title |
Extracted title |
content |
Markdown or HTML |
content_format |
markdown or html |
links |
Normalized extracted links |
video_metadata |
Present for supported video transcript extraction results |
blocked |
Likely anti-bot / verification / denied result |
llm_used |
Whether LLM enhancement was actually applied |
llm_error |
Why the LLM step degraded |
search_web
{
"query": "crawl4ai web scraping",
"engine": "auto",
"max_results": 10,
"lang": ""
}
| Parameter | Default | Description |
|---|---|---|
query |
(required) | Search query string |
engine |
auto |
Engine to use: auto, google, bing, duckduckgo, baidu |
max_results |
10 |
Maximum number of results |
lang |
"" |
Language hint (e.g. en, zh-CN) |
When engine="auto", the server tries engines in fallback order: DuckDuckGo → Bing → Google → Baidu. The first engine that returns results wins.
Search return shape
| Field | Meaning |
|---|---|
engine |
Which engine actually returned results |
query |
Original query |
results |
List of {title, url, snippet} |
total |
Number of results |
fallback_engines_tried |
Engines that failed before the successful one |
Anti-bot realism
The server already includes randomized browser behavior in code:
| Mechanism | Actual status |
|---|---|
| Random viewport | Yes |
| Random user agent mode | Yes, when explicit UA is not provided |
| Delay jitter | Yes |
override_navigator |
Yes |
simulate_user |
Yes, in stronger fallback mode |
| Proxy / cookies / persistent profile | Supported via env vars |
| Cloudflare bypass | Enhanced browser fingerprinting + configurable wait strategies |
Note: For overseas websites (Medium, ProductHunt, etc.), using a proxy is recommended. The server supports HTTP/HTTPS/SOCKS5 proxies via CRAWL4AI_MCP_PROXY environment variable.
Proxy input formats
CRAWL4AI_MCP_PROXY accepts all of these:
| Input | Interpreted as |
|---|---|
http://127.0.0.1:7890 |
HTTP proxy |
https://127.0.0.1:7890 |
HTTPS proxy |
socks5://127.0.0.1:7890 |
SOCKS5 proxy |
socket5://127.0.0.1:7890 |
Auto-normalized to socks5://... |
127.0.0.1:7890 |
Auto-normalized to http://127.0.0.1:7890 |
7890 |
Auto-normalized to http://127.0.0.1:7890 |
That means the README should not claim “perfect stealth”, but it can honestly claim human-like randomization and practical anti-bot knobs.
Quickstart
Conda
conda env create -f environment.yml
conda activate crawl4ai-mcp
python -m playwright install
crawl4ai-mcp
venv
python -m venv .venv
source .venv/bin/activate
python -m pip install -U pip
python -m pip install -e '.[dev]'
python -m playwright install
crawl4ai-mcp
MCP server config example
{
"mcpServers": {
"crawl4ai": {
"command": "crawl4ai-mcp",
"env": {
"CRAWL4AI_MCP_HEADLESS": "true",
"CRAWL4AI_MCP_PROXY": "127.0.0.1:7890",
"CRAWL4AI_MCP_NAVIGATION_TIMEOUT_MS": "30000",
"CRAWL4AI_MCP_WAIT_UNTIL": "load",
"OPENAI_BASE_URL": "https://your-openai-compatible-host",
"OPENAI_API_KEY": "your-api-key",
"OPENAI_MODEL": "your-model-name"
}
}
}
}
LLM-related env vars are optional. use_llm is still off by default at call time. If any LLM env is missing, invalid, or the model call fails, the server automatically falls back to non-LLM extraction.
Runtime configuration
| Env var | Purpose |
|---|---|
CRAWL4AI_MCP_HEADLESS |
Run browser headless |
CRAWL4AI_MCP_PROXY |
Upstream proxy, supports http://, https://, socks5://, host:port, and port-only |
CRAWL4AI_MCP_COOKIES_JSON |
Playwright storage state JSON |
CRAWL4AI_MCP_YTDLP_COOKIES_FROM_BROWSER |
Browser cookies source for video transcript extraction, e.g. chrome, firefox:default |
CRAWL4AI_MCP_YTDLP_COOKIEFILE |
Netscape cookies.txt path for yt-dlp video transcript extraction |
CRAWL4AI_MCP_USE_PERSISTENT_CONTEXT |
Reuse browser profile |
CRAWL4AI_MCP_USER_DATA_DIR |
Profile directory |
CRAWL4AI_MCP_NAVIGATION_TIMEOUT_MS |
Default max single navigation wait, default 30000 |
CRAWL4AI_MCP_WAIT_UNTIL |
Default page readiness strategy, default load |
OPENAI_BASE_URL |
OpenAI-compatible base URL |
OPENAI_API_KEY |
API key |
OPENAI_MODEL |
Model name |
One-shot CLI
This project now exposes a stateless one-shot CLI in addition to the MCP stdio server.
Fetch a single URL once and print JSON:
crawl4agent fetch "https://obsidian.md/help/cli" --format markdown
Search the web once and print JSON:
crawl4agent search "agent framework" --engine auto --max-results 5
Use proxy and browser cookies for video transcript extraction:
crawl4agent fetch "https://www.youtube.com/watch?v=OFfwN23hR8U" \
--proxy http://127.0.0.1:7890 \
--cookies-from-browser chrome
Run golden smoke once and print a JSON array:
crawl4agent smoke --out-dir ./_golden_outputs
The existing crawl4ai-mcp command remains the MCP stdio server entrypoint for MCP hosts.
Available help surfaces:
crawl4agent --help
crawl4agent fetch --help
crawl4agent search --help
crawl4agent smoke --help
Golden smoke regression
CRAWL4AI_MCP_SMOKE_DIR=./_golden_outputs .venv/bin/python -m crawl4ai_mcp.smoke_golden
For overseas video URLs, a local proxy is often needed:
CRAWL4AI_MCP_PROXY=http://127.0.0.1:7890 \
CRAWL4AI_MCP_SMOKE_DIR=./_golden_outputs \
.venv/bin/python -m crawl4ai_mcp.smoke_golden
This writes full markdown outputs to _golden_outputs/ so you can inspect extraction quality page by page.
The golden set now includes the earlier baseline URLs plus ainew.me, openclaw, watcha, producthunt, mydrivers, caihongtu, openrouter, mobile Douban, and video pages from YouTube / Bilibili. For sites outside mainland China, proxy-based verification is recommended.
Some overseas sites may still return Cloudflare or similar verification pages even when a proxy is configured. In those cases the server now marks them with blocked=true. The recommended path is: better proxy quality, valid cookies, or a persistent browser profile after manual verification.
For some video golden URLs, subtitle extraction may require login. If yt-dlp reports login-required subtitles, configure either CRAWL4AI_MCP_YTDLP_COOKIES_FROM_BROWSER or CRAWL4AI_MCP_YTDLP_COOKIEFILE before running golden smoke.
Prior art
- Crawl4AI: https://github.com/unclecode/crawl4ai
- mcp-crawl4ai-rag: https://github.com/coleam00/mcp-crawl4ai-rag
- weidwonder/crawl4ai-mcp-server: https://github.com/weidwonder/crawl4ai-mcp-server
- WaterCrawl: https://github.com/watercrawl/WaterCrawl
- teracrawl: https://github.com/BrowserCash/teracrawl
License
This project is licensed under AGPL-3.0-or-later.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。