crawl4ai-cloud-sdk 0.6.1

Lightweight cloud SDK for Crawl4AI - mirrors the OSS API

Crawl4AI Cloud SDK

The fastest way to turn any URL into markdown, screenshots, structured data, or a full site crawl.

Install

pip install crawl4ai-cloud-sdk

Get your API key at api.crawl4ai.com.

Quick Start

import asyncio
from crawl4ai_cloud import AsyncWebCrawler

async def main():
    async with AsyncWebCrawler(api_key="sk_live_...") as crawler:

        # Get clean markdown from any page
        md = await crawler.markdown("https://example.com")
        print(md.markdown)

        # Take a full-page screenshot
        ss = await crawler.screenshot("https://example.com")
        # ss.screenshot is base64-encoded PNG

        # Extract structured data with natural language
        data = await crawler.extract(
            "https://news.ycombinator.com",
            query="get each story title, URL, and points",
        )
        print(data.data)  # list of dicts

        # Discover all URLs on a domain
        sitemap = await crawler.map("https://docs.python.org")
        for u in sitemap.urls[:10]:
            print(u.url)

        # Crawl an entire site (async, returns job_id)
        job = await crawler.crawl_site(
            "https://docs.example.com",
            max_pages=50,
            wait=True,
        )
        print(f"Done: {job.discovered_urls} pages crawled")

asyncio.run(main())

Wrapper API Reference

Method	What it does	Endpoint
markdown(url)	Returns clean markdown (with optional fit/pruning)	POST /v1/markdown
screenshot(url)	Returns base64 screenshot (PNG) and optional PDF	POST /v1/screenshot
extract(url, query=...)	Extracts structured data (auto/LLM/CSS schema)	POST /v1/extract
map(url)	Simple URL discovery on a domain (always sync)	POST /v1/map
scan(url, criteria=...)	AI-assisted URL discovery with plain-English criteria + map/deep routing	POST /v1/scan
crawl_site(url, criteria=..., extract=...)	AI-assisted full site crawl -- LLM generates scan config + extraction schema	POST /v1/crawl/site
enrich(query=…)	Build a data table from a brief, list of entities, or list of URLs (multi-phase)	POST /v1/enrich/async

Each method returns a typed response object (MarkdownResponse, ScreenshotResponse, ExtractResponse, MapResponse, ScanResult, SiteCrawlResponse, EnrichJobStatus) with relevant status and data fields.

AI-assisted flows (v0.4.0)

Pass a plain-English criteria and let the backend LLM pick scan mode, URL patterns, filters, and scorers. Pair with extract on crawl_site() to also auto-generate a CSS extraction schema from a sample URL. The generated config is echoed back so you can see and reuse it.

async with AsyncWebCrawler(api_key="sk_live_...") as crawler:

    # AI-assisted scan — LLM picks map vs deep + generates patterns/query/threshold
    result = await crawler.scan(
        "https://docs.crawl4ai.com",
        criteria="API reference and core docs pages",
        max_urls=50,
    )
    print(f"Mode: {result.mode_used}")                  # "map" or "deep"
    print(f"Found: {result.total_urls} URLs")
    if result.generated_config:
        print(f"AI: {result.generated_config.reasoning}")

    # Explicit deep scan with async polling
    job = await crawler.scan(
        "https://directory.example.com",
        criteria="company profile pages",
        scan={"mode": "deep", "max_depth": 3},
        wait=True,                                       # block until done
        poll_interval=3.0,
    )

    # Flagship: crawl whole site + auto-extract structured data
    job = await crawler.crawl_site(
        "https://books.toscrape.com",
        criteria="all book listing pages",
        max_pages=50,
        strategy="http",
        extract={
            "query": "book title, price, rating",
            "json_example": {"title": "...", "price": "£0.00", "rating": 0},
            "method": "auto",                            # picks CSS schema vs LLM
        },
        include=["links"],                               # drop markdown — extract-only
    )
    print(f"Generated schema: {bool(job.schema_used)}")
    print(f"Method: {job.extraction_method_used}")      # "css_schema" or "llm"

    # Unified polling — one endpoint for scan + crawl phases
    while True:
        status = await crawler.get_site_crawl_job(job.job_id)
        print(f"{status.phase}: {status.progress.urls_crawled}/{status.progress.total}")
        if status.is_complete:
            print(f"Download: {status.download_url}")
            break
        await asyncio.sleep(3)

Config objects (optional — both scan and extract accept plain dicts or typed dataclasses):

from crawl4ai_cloud import SiteScanConfig, SiteExtractConfig

scan_cfg = SiteScanConfig(
    mode="auto",                                         # "auto" | "map" | "deep"
    patterns=["*/docs/*", "*/guide/*"],
    scorers={"keywords": ["auth", "oauth"], "optimal_depth": 2},
    max_depth=3,
)

extract_cfg = SiteExtractConfig(
    query="book title, price, rating",
    json_example={"title": "...", "price": "£0.00", "rating": 0},
    method="auto",
)

job = await crawler.crawl_site(
    "https://books.toscrape.com",
    criteria="book listings",
    scan=scan_cfg,
    extract=extract_cfg,
)

Drop markdown with include: if you pass include=["links", "media"] without "markdown", the worker force-strips markdown from every result -- saves bandwidth for extract-only crawls.

Enrich v2 (v0.6.0)

Multi-phase enrichment. Give a brief, a list of entities, or a list of URLs and get back a structured table with per-field provenance, certainty, and disputed-value markers.

The job walks through phases: queued → planning → plan_ready → resolving_urls → urls_ready → extracting → merging → completed. Defaults auto_confirm_plan=True, auto_confirm_urls=True make it run straight through (one-shot). Set either to False for human-in-loop review and resume via resume_enrich_job(...).

async with AsyncWebCrawler(api_key="sk_live_...") as crawler:

    # 1. Agent one-shot — give a brief, get a table back
    result = await crawler.enrich(
        query="licensed nurseries in North York Toronto with extended hours",
        country="ca",
        top_k_per_entity=3,
    )
    for row in result.rows:
        print(row.input_key, row.fields)
        # certainty + sources are per-field
        for f, c in row.certainty.items():
            print(f"  {f}: {c:.2f}  (from {row.sources[f]['url']})")
    print(f"Crawls: {result.usage.crawls}, Searches: {result.usage.searches}")
    print(f"LLM totals: {result.usage.llm_totals}")

    # 2. Pre-resolved URLs — skip planning + URL resolution
    result = await crawler.enrich(
        urls=["https://example.com/a", "https://example.com/b"],
        features=["price", "hours"],   # string shortcut: same as [{"name": "price"}, ...]
    )

    # 3. Human review flow — pause for editing the plan
    job = await crawler.enrich(
        query="best Italian restaurants in Brooklyn",
        country="us",
        auto_confirm_plan=False,
        auto_confirm_urls=False,
        wait=False,
    )

    # Wait for the planning phase to land
    job = await crawler.wait_enrich_job(job.job_id, until="plan_ready")
    print(job.plan.entities, job.plan.features)

    # Edit and resume — the server applies your edits then advances
    await crawler.resume_enrich_job(
        job.job_id,
        entities=[{"name": "Lucali"}, {"name": "Roberta's"}],
        features=[{"name": "address"}, {"name": "hours"}],
    )

    # Wait again for the URL-resolution pause, then resume to completion
    job = await crawler.wait_enrich_job(job.job_id, until="urls_ready")
    await crawler.resume_enrich_job(job.job_id)   # accept server's URL picks
    final = await crawler.wait_enrich_job(job.job_id)

    # 4. Live progress via SSE
    async for event in crawler.stream_enrich_job(job.job_id):
        if event.type == "phase":  print("→", event.status)
        elif event.type == "row":  print("✓", event.row.input_key)
        elif event.type == "complete": break

    # Job management
    jobs = await crawler.list_enrich_jobs(limit=5)
    await crawler.cancel_enrich_job(job.job_id)

Vocabulary:

• Entity — one row identifier (e.g. "Franklin Barbecue").
• Criterion — a search-side filter when finding URLs per entity ({"text": "Austin TX", "kind": "location"}).
• Feature — one extraction column read off each crawled page.

Per-purpose usage is reported in result.usage.llm_tokens_by_purpose with five buckets: plan_intent, url_plan, paywall_classify, extract, merge_tiebreak (only buckets that ran appear).

Async / Batch

Every wrapper method has a _many variant for processing multiple URLs as an async job.

async with AsyncWebCrawler(api_key="sk_live_...") as crawler:

    # Batch markdown (fire-and-forget)
    job = await crawler.markdown_many(
        ["https://a.com", "https://b.com", "https://c.com"],
    )
    print(f"Job {job.job_id} started, {job.urls_count} URLs queued")

    # Batch markdown (wait for results)
    job = await crawler.markdown_many(urls, wait=True, timeout=120)

    # Batch screenshots
    job = await crawler.screenshot_many(urls, full_page=True, wait=True)

    # Batch extraction (note: method must be "llm" or "schema", not "auto")
    job = await crawler.extract_many(
        urls, method="llm", query="get product name and price", wait=True,
    )

    # Site crawl is always async — prefer criteria + extract over legacy discovery flag
    site = await crawler.crawl_site(
        "https://docs.example.com",
        criteria="all API reference pages",
        max_pages=100,
        wait=True,
    )

Job Management

Each wrapper namespace has its own job management methods.

# Markdown jobs
job   = await crawler.get_markdown_job(job_id)
jobs  = await crawler.list_markdown_jobs(status="completed", limit=10)
await crawler.cancel_markdown_job(job_id)

# Screenshot jobs
job   = await crawler.get_screenshot_job(job_id)
jobs  = await crawler.list_screenshot_jobs()
await crawler.cancel_screenshot_job(job_id)

# Extract jobs
job   = await crawler.get_extract_job(job_id)
jobs  = await crawler.list_extract_jobs()
await crawler.cancel_extract_job(job_id)

# Scan jobs (AI-assisted deep scans)
job   = await crawler.get_scan_job(job_id)              # unified status + URLs-so-far
await crawler.cancel_scan_job(job_id)                   # preserves partial results

# Site crawl jobs (unified scan + crawl polling)
job   = await crawler.get_site_crawl_job(job_id)        # phase: scan|crawl|done
# Cancel delegates to the underlying deep crawl job:
await crawler.cancel_deep_crawl(job_id)

# Core crawl jobs (from run_many / deep_crawl)
job   = await crawler.get_job(job_id)
jobs  = await crawler.list_jobs(status="running")
await crawler.cancel_job(job_id)
url   = await crawler.download_url(job_id)  # presigned S3 ZIP

Power User: Config Passthrough

All wrapper methods accept crawler_config and browser_config dicts for full control. These are the same fields you would pass to the core /v1/crawl endpoint.

md = await crawler.markdown(
    "https://example.com",
    strategy="browser",
    fit=True,
    include=["links", "media", "tables"],
    crawler_config={
        "css_selector": "article",
        "exclude_external_links": True,
        "wait_for": ".content-loaded",
        "js_code": "window.scrollTo(0, document.body.scrollHeight)",
    },
    browser_config={
        "viewport_width": 1920,
        "viewport_height": 1080,
        "headers": {"Accept-Language": "en-US"},
    },
    proxy="residential",
)

Works the same way for screenshot(), extract(), map(), and crawl_site().

Full Power Mode

For advanced use cases where you need full control over the crawl pipeline, the core methods give you direct access to the /v1/crawl endpoint with every configuration option.

Single URL

from crawl4ai_cloud import CrawlerRunConfig, BrowserConfig

config = CrawlerRunConfig(
    screenshot=True,
    word_count_threshold=10,
    exclude_external_links=True,
    process_iframes=True,
    css_selector="article",
)
browser_config = BrowserConfig(
    viewport_width=1920,
    viewport_height=1080,
)

result = await crawler.run(
    "https://example.com",
    config=config,
    browser_config=browser_config,
    proxy="datacenter",
)
print(result.markdown.raw_markdown)
print(result.screenshot)  # base64

Batch Crawl

job = await crawler.run_many(
    ["https://a.com", "https://b.com"],
    config=config,
    wait=True,
    priority=1,
)
# Results available via download
url = await crawler.download_url(job.id)

Deep Crawl

result = await crawler.deep_crawl(
    "https://docs.example.com",
    strategy="bfs",       # bfs, dfs, best_first, map
    max_depth=3,
    max_urls=100,
    include_patterns=["docs", "api"],
    exclude_patterns=["download"],
    wait=True,
)

Domain Scan

scan = await crawler.scan("https://example.com", mode="deep", max_urls=200)
for url_info in scan.urls:
    print(f"{url_info.url} (score: {url_info.relevance_score})")

Full reference: Cloud API Docs

Configuration

CrawlerRunConfig

Controls what gets extracted and how pages are processed.

from crawl4ai_cloud import CrawlerRunConfig

config = CrawlerRunConfig(
    css_selector="main",          # target specific elements
    excluded_tags=["nav", "footer"],
    word_count_threshold=10,
    screenshot=True,
    wait_for=".loaded",           # wait for CSS selector
    js_code="document.querySelector('.show-more').click()",
    magic=True,                   # anti-bot mode
)

BrowserConfig

Controls the browser environment.

from crawl4ai_cloud import BrowserConfig

browser = BrowserConfig(
    viewport_width=1920,
    viewport_height=1080,
    user_agent="MyBot/1.0",
    headers={"Authorization": "Bearer token"},
    cookies=[{"name": "session", "value": "abc", "domain": "example.com"}],
    profile_id="my-saved-profile",  # cloud browser profile
)

ProxyConfig

from crawl4ai_cloud import ProxyConfig

# Shorthand (works on all methods)
result = await crawler.markdown(url, proxy="datacenter")
result = await crawler.markdown(url, proxy="residential")

# Full config
proxy = ProxyConfig(mode="residential", country="US", sticky_session=True)
result = await crawler.markdown(url, proxy=proxy)

Proxy modes: "none" (direct, 1x credits), "datacenter" (fast, 2x), "residential" (premium, 5x), "auto" (smart selection).

Environment Variables

export CRAWL4AI_API_KEY=sk_live_...

# API key auto-loaded from environment
async with AsyncWebCrawler() as crawler:
    md = await crawler.markdown("https://example.com")

Error Handling

from crawl4ai_cloud import (
    CloudError,
    AuthenticationError,
    RateLimitError,
    QuotaExceededError,
    NotFoundError,
    ValidationError,
    TimeoutError,
    ServerError,
)

try:
    result = await crawler.markdown(url)
except AuthenticationError:
    print("Invalid API key")
except RateLimitError as e:
    print(f"Rate limited. Retry after {e.retry_after}s")
except QuotaExceededError as e:
    print(f"Quota exceeded ({e.quota_type})")
except TimeoutError:
    print("Request timed out")
except ValidationError:
    print("Invalid request parameters")
except ServerError:
    print("Server error, try again later")
except CloudError as e:
    print(f"[{e.status_code}] {e.message}")

Claude Code Plugin

Use Crawl4AI directly inside Claude Code with 9 built-in tools.

/plugin marketplace add unclecode/crawl4ai-cloud-sdk
/plugin install crawl4ai@crawl4ai-claude-plugins

See plugin README for details.

Links

• Cloud Dashboard - Sign up and manage your API key
• Cloud API Docs - Full API reference
• PyPI - Package page
• GitHub - Source code
• OSS Crawl4AI - Self-hosted option
• Discord - Community and support

License

Apache 2.0