CrawlEyes MCP Server

CrawlEyes — Web Scraping & Search Toolkit for AI Agents

English | 中文

CrawlEyes gives AI agents reliable full-text extraction (web_extract) and robust search (web_search) backends — the "eyes" that let agents see and read the web. Built and tested against Hermes Agent.

Also ships as a standard MCP server, so any MCP client (Claude Desktop, Cursor, other agents) can reuse the same search + extraction capabilities.

Why CrawlEyes?

Most agent toolkits cover *one* slice of the pipeline. CrawlEyes is the rare all-in-one that you can actually run behind the Great Firewall without external accounts.

Zero API keys. Zero external accounts. One command. CrawlEyes is the only toolkit in this space that combines search + extraction + semantic reranking + MCP in a single, China-friendly, self-hosted package.

Features

Project layout

plugins/searxng-tavily/   Hermes web-search provider plugin (SearXNG → Tavily keyless fallback)

• three-state circuit breaker + shared SQLite cache

scripts/ crawl4ai_cli.py Universal scraping CLI (URL → Markdown), with denoise/retry/session/BM25 crawl_search_standalone.py Standalone search (SearXNG → Tavily) + optional semantic rerank. No Hermes dependency — usable anywhere, powers the MCP server. mcp_crawl_server.py Standard MCP server exposing search + extract + deep_research + sitemap (stdio) single_env_check.py Verify crawl4ai provider registered+available+extracts (one profile) verify_searxng_tavily.py Verify searxng-tavily provider: normal path + forced fallback agent_link_check.py Verify full agent tool chain: web_search_tool dispatch + logs

Quick start

1. Install Crawl4AI (China-friendly mirrors)

python3 -m venv .venv

Use Tsinghua PyPI mirror for speed (or any mirror you prefer)

.venv/bin/pip install -i https://pypi.tuna.tsinghua.edu.cn/simple crawl4ai

Playwright browser kernel — use npmmirror binary mirror if cdn.playwright.dev is blocked

PLAYWRIGHT_DOWNLOAD_HOST=https://registry.npmmirror.com/-/binary/playwright \ .venv/bin/python -m playwright install chromium .venv/bin/crawl4ai-setup

2. Scrape a page

.venv/bin/python scripts/crawl4ai_cli.py https://example.com          # stdout Markdown

.venv/bin/python scripts/crawl4ai_cli.py https://example.com -o out.md # to file .venv/bin/python scripts/crawl4ai_cli.py URL --text --max-words 5000 # plain text, truncated

Multi-format extraction (markdown|fit|raw|markdown_with_citations)

.venv/bin/python scripts/crawl4ai_cli.py URL --format raw # unfiltered source markdown .venv/bin/python scripts/crawl4ai_cli.py URL --format markdown_with_citations # + source URLs

Respect robots.txt (opt-in, default off)

.venv/bin/python scripts/crawl4ai_cli.py URL --respect-robots

Denoise nav/ads + retry 3x + reuse session across scrapes

.venv/bin/python scripts/crawl4ai_cli.py URL --noise-filter --retry 3 --session s1

3. Use the search + rerank (standalone, no Hermes)

# Optional: local semantic rerank of results (fastembed + bge-small-zh, auto-downloaded)

.venv/bin/pip install -i https://pypi.tuna.tsinghua.edu.cn/simple fastembed

SearXNG first, Tavily keyless fallback, then rerank

SEARXNG_URL=https://your-searxng .venv/bin/python -c " import sys; sys.path.insert(0, 'scripts') from crawl_search_standalone import CrawlSearch r = CrawlSearch(rerank=True).search('your query') print(r'data')"

China-network note: the embedding model downloads from HuggingFace, which is blocked on mainland networks. Set HF_ENDPOINT=https://hf-mirror.com and HF_HUB_DISABLE_XET=1 (hf-mirror doesn't support the xet protocol and returns 401 without this).

4. Run as an MCP server (any client)

# Any MCP client can connect via stdio (default):

.venv/bin/python scripts/mcp_crawl_server.py

Exposes tools:

search(query, limit) - SearXNG → Tavily keyless, rerank, retry+rate-limit

extract(url, max_words, format) - markdown|fit|raw|markdown_with_citations

deep_research(topic, num_questions) - multi-round cited report

sitemap(origin, max_urls) - URL map from sitemap.xml / robots.txt

Or serve over HTTP (streamable-http) for remote clients:

.venv/bin/python -m crawleyes.mcp_crawl_server --transport http --port 8765 --host 127.0.0.1

→ clients connect to http://127.0.0.1:8765/mcp

(host/port configurable; default 127.0.0.1:8765)

For Hermes specifically, add to config.yaml:

mcp_servers:

crawl: command: "/path/to/crawl/.venv/bin/python" args: ["/path/to/crawl/scripts/mcp_crawl_server.py"] timeout: 90 connect_timeout: 60

4b. Firecrawl-compatible /scrape endpoint

Already using Firecrawl's Python SDK? Point it at CrawlEyes and keep your code:

.venv/bin/python -m crawleyes.firecrawl_api --port 8899 --host 127.0.0.1

POST /v2/scrape → { success, data: { markdown, metadata } }

GET /healthz → health check

from firecrawl import Firecrawl

fc = Firecrawl(api_url="http://127.0.0.1:8899", api_key="ignored") doc = fc.scrape(url="https://example.com") # → { markdown, metadata }

This is a pragmatic subset of the Firecrawl API — the core /scrape contract (success + data.markdown + data.metadata), backed by CrawlEyes' own extraction engine. It does not implement Firecrawl's async /crawl queue, /search, or /map — see the design notes for the rationale.

5. Install the search plugin (Hermes)

Copy plugins/searxng-tavily/ into a Hermes plugins dir, then:

hermes plugins enable web/searxng-tavily

hermes config set web.search_backend searxng-tavily

Set SEARXNG_URL in your Hermes profile .env to point at your SearXNG instance. If unset or unreachable, the provider automatically falls back to the Tavily keyless API (no API key required).

Note: the plugin only takes effect for newly started agent sessions.

6. Verify

# Requires the Hermes source tree + its venv

venv/bin/python scripts/verify_searxng_tavily.py $HERMES_HOME venv/bin/python scripts/agent_link_check.py $HERMES_HOME

Design notes

• Layered composition: no single tool covers everything. Crawl4AI handles extraction; SearXNG + Tavily cover search; each layer has a tested fallback.
• Tavily keyless works with zero configuration and no account — a cheap resilience net for the whole search path.
• Circuit breaker is SearXNG-only: a Tavily fallback success does *not* reset the breaker (otherwise it would never trip). record_success() is only called when SearXNG itself succeeds.
• Shared SQLite cache lives in the *real* user home (via pwd.getpwuid, not $HOME — which Hermes profiles override), so all profiles share one cache. WAL + 5s timeout + try/except degrade-to-no-cache under concurrency.
• Semantic rerank is cheap: fastembed (ONNX) avoids the ~2GB torch dependency; model loads in ~0.6s once cached, embeddings in ~50ms.
• MCP server is standalone: it does *not* import Hermes internals, so it runs on any Python 3.12 env and serves any MCP client.
• MCP transport is dual: stdio (default, standard MCP clients) or streamable-http (--transport http), so a single codebase serves bot