CrawlEyes — Web Scraping & Search Toolkit for AI Agents

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.

	Typical agent toolkit	CrawlEyes
🔍 Search	API key required, often blocked in CN	✅ SearXNG (self-hosted) + Tavily keyless fallback — zero config, zero key
📄 Extraction	Separate scraper, or Firecrawl SaaS	✅ Built-in Crawl4AI full-text extract, ~89% noise removal
🧠 Semantic rerank	Rarely included	✅ Local fastembed rerank — no torch, ~50MB model
🔌 MCP server	Often missing	✅ Standard MCP tools (search + extract + deep_research + sitemap), any client
🌐 China-friendly	Mostly English/GFW-blocked	✅ Tested on a real mainland China server (baidu + yandex)

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

Capability	Where	Why it matters
Full-text extraction	scripts/crawl4ai_cli.py	Headless-browser scraping → clean Markdown; handles ~80% of JS/dynamic/UA-blocked pages
Content denoising (P1)	crawl4ai_cli.py --noise-filter	Prunes nav/ads/comments via Crawl4AI's PruningContentFilter — measured 24.6k→2.8k chars (~89% noise removed) on a typical article
Retry with backoff (P3)	crawl4ai_cli.py --retry N	Exponential backoff (1s/2s/4s) on transient failures
Browser session reuse (P4)	crawl4ai_cli.py --session NAME	Reuses the browser context across scrapes in one process — no cold-start per URL
Keyword-focused extraction	crawl4ai_cli.py --bm25 KEYWORD	Keeps only paragraphs relevant to a keyword (experimental — BM25 is English-centric; works best on English docs)
Search (primary)	SearXNG (self-hosted meta-search)	Privacy-friendly search aggregator
Search (fallback)	Tavily keyless API	Zero-config, no-key fallback when SearXNG is down/empty
Search orchestration	plugins/searxng-tavily/	Hermes plugin provider: SearXNG first → auto-fallback to Tavily keyless; three-state circuit breaker (3 fails → 60s cooldown → half-open) + shared SQLite cache (TTL 3600s)
Semantic reranking (P2)	scripts/crawl_search_standalone.py	Local embedding rerank of search results with fastembed + BAAI/bge-small-zh-v1.5 (512-dim, no torch dependency, ~50MB, cached) — puts relevant results first. Measured: crawler-relevant items 0.817/0.732 float to top, irrelevant 0.302/0.139 sink
MCP server (P5)	scripts/mcp_crawl_server.py	Exposes search + extract + deep_research + sitemap as standard MCP tools (stdio default, or streamable-http for remote clients). Works in any MCP client, no Hermes dependency. Extracted content is sanitized against prompt-injection (strips invisible chars + prompt-hijack lines). Unified rate limiting + exponential backoff guard every tool (sliding window, per-tool cost) so concurrent agent calls can't hammer downstream services
Sitemap discovery (P1)	crawleyes/sitemap.py	sitemap(origin) → parses sitemap.xml (plain / gzip / index-recursion) with robots.txt fallback, returns a deduped URL map. Zero-key way to discover a site's URL surface for whole-site fetch or deep-research seeding
Multi-format extract (P0)	extract(..., format=)	markdown (default) / fit (denoised) / raw (unfiltered) / markdown_with_citations — pick the level of cleanup you need
RAG-ready interfaces	crawleyes/rag.py	One-liners markdown(url) / search_markdown(query) → clean, sanitized, LLM-ready Markdown for RAG corpora
Deep research	crawleyes/deep_research.py	deep_research(topic) → decomposes topic into sub-questions → searches → extracts → synthesizes a cited Markdown report. Optional LLM (any OpenAI-compatible endpoint); degrades to evidence-aggregate mode without one
Verification	scripts/	Clean subprocess scripts to verify each backend end-to-end per Hermes profile

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']['web'])"

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 both local process and remote HTTP clients.
• Unified rate limiting is layered: MCP tools and deep-research's internal search/extract all share one sliding-window limiter (per-tool cost), so concurrent agent fan-out can't hammer SearXNG/Tavily/Crawl4AI even through multi-round deep research.

Credits & inspiration

This project builds on a set of excellent open-source tools. All code here is an independent implementation (no copied code), but the ideas and interfaces are drawn from the following projects — full credit to their authors:

Feature in this repo	Inspired by	License
Extraction engine (Crawl4AI wrapper)	Crawl4AI — direct dependency	Apache-2.0
Content denoising (P1)	Readability, GeneralNewsExtractor (idea)	Apache-2.0 / MIT
Semantic reranking (P2)	Vane, Perplexica (idea)	MIT / MIT
Retry with backoff (P3)	Crawlee (idea)	Apache-2.0
Browser session reuse (P4)	camoufox (idea)	MIT
MCP server (P5)	playwright-mcp, exa-mcp-server (idea)	Apache-2.0 / MIT
Search orchestration / fallback	SearXNG — self-hosted (official Docker image, no source modification), accessed via HTTP API only · Tavily keyless	AGPL-3.0 (server software, not linked/embedded) / proprietary API

Design independence: the implementations here are written from scratch — we studied the above projects' approaches (denoising thresholds, rerank pipelines, backoff strategies, MCP tool patterns) but did not copy their source code. Dependencies are declared in requirements.txt. If you believe any attribution is missing or incorrect, please open an issue.

Compliance

CrawlEyes is a general-purpose fetch toolkit for legitimate research and personal use. It deliberately does not include proxy pools, fingerprint rotation, or CAPTCHA-solving (anti-scraping evasion) — those are out of scope.

Robots.txt is opt-in (default off): pass respect_robots=True to extract / markdown (or --respect-robots on the CLI) to check each target's robots.txt (RFC 9309) and refuse URLs it explicitly disallows. It's default-off so legitimate scraping isn't silently blocked by aggressive or broken robots rules — compliance is the caller's informed choice per use case. Always review each site's terms of service before scraping at scale.

License

MIT — see LICENSE.