silkworm-rs

Async-first web scraping framework built on wreq (HTTP with browser impersonation) and scraper-rs (fast HTML parsing). Silkworm gives you a minimal Spider/Request/Response model, middlewares, and pipelines so you can script quick scrapes or build larger crawlers without boilerplate.

📖 Documentation: https://bitingsnakes.github.io/silkworm/

Features

• Async engine with configurable concurrency, priority-aware queueing, bounded, deadlock-free backpressure (defaults to concurrency * 10; hard for seeding, soft when several callbacks produce requests at once), and per-request timeouts.
• wreq-powered HTTP client: browser impersonation, redirect following with loop detection, query merging, and proxy support via request.meta["proxy"].
• Optional OnionLink client integration for scraping Tor v3 .onion sites without routing through wreq.
• Optional Servo rendering via ServoFetchClient for JavaScript-rendered pages without changing the default HTTP client.
• Typed async spiders with a push-style callback API: await self.emit(item) streams items to pipelines and await self.follow(...) / await response.follow(href) schedule requests, both with backpressure; HTMLResponse ships selector helpers.
• Optional declarative extraction with compiled Item, Text, and Attr field plans while keeping the callback API available.
• HTML-to-Markdown conversion via fast-h2m, including rich full, lean minimal, and streaming modes.
• Middlewares: User-Agent rotation/default, proxy rotation, cookie jars with save/load, retries of error statuses and transient network failures with exponential backoff, per-host AutoThrottle, robots.txt enforcement (Disallow and Crawl-delay), flexible delays, SkipNonHTMLMiddleware to drop non-HTML callbacks, and CloudflareCrawlMiddleware for Browser Rendering crawl jobs.
• Pipelines: JSON Lines, SQLite, XML (nested data preserved), and CSV (flattens dicts and lists) out of the box, plus schema validation with ValidationPipeline.
• Production controls: a CrawlResult with a failure policy (error rate, minimum items, drop rate) so broken spiders fail loudly; stop limits (max_items, max_requests, max_depth, max_duration, max_errors); allowed_domains; per-domain concurrency; graceful SIGINT/SIGTERM shutdown; response size limits; pause/resume with job_dir; an on-disk HTTP cache; Prometheus metrics; and layered settings (environment, custom_settings, CLI).
• A silkworm command line (crawl, parse, fetch) and silkworm.testing helpers for testing callbacks offline.
• Structured logging via the standard library (SILKWORM_LOG_LEVEL=DEBUG), plus periodic/final crawl statistics with per-status, per-domain, and per-error breakdowns.

Installation

From PyPI with pip:

pip install silkworm-rs

From PyPI with uv (recommended for faster installs):

uv pip install silkworm-rs
# or if using uv's project management:
uv add silkworm-rs

From source:

uv venv  # install uv from https://docs.astral.sh/uv/getting-started/ if needed
source .venv/bin/activate  # Windows: .venv\Scripts\activate
uv pip install -e .

Targets Python 3.13+; dependencies are pinned in pyproject.toml.

Quick start

Define a spider by subclassing Spider and implementing an async parse that reports items with await self.emit(...) and schedules follow-up pages with await response.follow(...). This example writes quotes to data/quotes.jl and enables basic user agent, retry, and non-HTML filtering middlewares.

from silkworm import HTMLResponse, Response, Spider, run_spider
from silkworm.middlewares import (
    RetryMiddleware,
    SkipNonHTMLMiddleware,
    UserAgentMiddleware,
)
from silkworm.pipelines import JsonLinesPipeline


class QuotesSpider(Spider):
    name = "quotes"
    start_urls = ("https://quotes.toscrape.com/",)

    async def parse(self, response: Response) -> None:
        if not isinstance(response, HTMLResponse):
            return

        html = response
        for quote in await html.select(".quote"):
            text_el = await quote.select_first(".text")
            author_el = await quote.select_first(".author")
            if text_el is None or author_el is None:
                continue
            tags = await quote.select(".tag")
            await self.emit(
                {
                    "text": text_el.text,
                    "author": author_el.text,
                    "tags": [t.text for t in tags],
                }
            )

        if next_link := await html.select_first("li.next > a"):
            if href := next_link.attr("href"):
                await html.follow(href, callback=self.parse)


if __name__ == "__main__":
    run_spider(
        QuotesSpider,
        request_middlewares=[UserAgentMiddleware()],
        response_middlewares=[
            SkipNonHTMLMiddleware(),
            RetryMiddleware(max_times=3, sleep_http_codes=[429, 503]),
        ],
        item_pipelines=[JsonLinesPipeline("data/quotes.jl")],
        concurrency=16,
        request_timeout=10,
        log_stats_interval=30,
    )

Upgrading from 0.10

Silkworm 0.11 replaced yield/return-based callbacks with emit/follow. Callbacks, errbacks and start_requests() are now async functions returning None: turn yield item into await self.emit(item), yield request into await self.follow(request), and yield response.follow(href) into await response.follow(href). Legacy generator callbacks fail with a SpiderError that explains the fix. See the migration table.

Upgrading to 0.12

• Runners and Engine.run() return a CrawlResult instead of None.
• The default deduplication key is the request fingerprint (method, canonical URL with params, body) instead of the raw URL, so ?a=1&b=2 and ?b=2&a=1 are one page while POSTs with different bodies are not.
• items_scraped counts items that passed every pipeline; pipelines can raise DropItem to discard items (counted as items_dropped).
• Responses carry the exact downloaded bytes (earlier versions re-encoded bodies as UTF-8 text, corrupting binary files and non-UTF-8 pages), and bodies over max_response_size_bytes (default 50 MB) fail with ResponseTooLargeError.
• Timeouts and connection failures raise HttpTimeoutError/HttpConnectionError (both HttpError subclasses), which RetryMiddleware now retries.
• Requests record their link depth in meta["depth"], and the new counters' names (such as retries) are reserved in Spider.stats_payload.
• The sync runners stop gracefully on SIGINT/SIGTERM (handle_signals=False opts out).
• Requests time out after 60 seconds by default (request_timeout, also the HttpClient(timeout=...) default); pass None to restore the previous unlimited behaviour.

Production crawling

Declare what a successful crawl means, stop safely, stay polite, and resume after interruptions:

from silkworm import run_spider
from silkworm.middlewares import AutoThrottleMiddleware, RetryMiddleware, RobotsTxtMiddleware
from silkworm.pipelines import JsonLinesPipeline, ValidationPipeline

throttle = AutoThrottleMiddleware(start_delay=0.5, max_delay=30)
result = run_spider(
    QuotesSpider,
    request_middlewares=[RobotsTxtMiddleware(), throttle],
    response_middlewares=[throttle, RetryMiddleware(max_times=3)],
    item_pipelines=[ValidationPipeline(Quote), JsonLinesPipeline("data/quotes.jl")],
    request_timeout=30,
    concurrency_per_domain=4,
    max_depth=5,
    max_error_rate=0.05,  # raise CrawlFailedError above 5% failed requests
    min_items=50,  # ...or when fewer than 50 valid items were scraped
    job_dir="state/quotes",  # Ctrl+C, then run again to resume
    metrics_port=9410,  # Prometheus metrics at http://127.0.0.1:9410/metrics
)
print(result.close_reason, result.items_scraped, result.error_rate)

Or from the command line, with settings from -s, SILKWORM_* environment variables, or Spider.custom_settings:

silkworm crawl examples/quotes_spider.py -o data/quotes.jl -s max_items=100 --job-dir state/quotes
silkworm parse https://quotes.toscrape.com/ --spider examples/quotes_spider.py

silkworm crawl exits with status 1 when the failure policy is violated, so cron jobs and CI notice broken spiders. See the Production Crawling guide and the CLI reference.

Declarative extraction

Use silkworm.declarative when the extraction is regular and keep ordinary callbacks for pagination or site-specific behavior:

from silkworm import HTMLResponse, Response, Spider
from silkworm.declarative import Attr, Item, Text


def parse_price(value: str) -> float:
    return float(value.removeprefix("$").strip())


class Product(Item):
    __selector__ = ".product"

    title: str = Text("h2", strip=True)
    price: float = Text(".price", transform=parse_price)
    url: str = Attr("a", "href", absolute=True)
    image: str | None = Attr("img", "src", absolute=True)
    tags: list[str] = Text(".tag")


class ProductsSpider(Spider):
    start_urls = ("https://shop.example.com/products/",)

    async def parse(self, response: Response) -> None:
        if not isinstance(response, HTMLResponse):
            return

        async for product in Product.extract(response):
            # Existing pipelines consume JSON-compatible values.
            await self.emit(product.to_dict())

The annotation controls selector cardinality:

Fields support these options:

• default=... supplies the final scalar value when an element or attribute is missing.
• transform=... applies a synchronous conversion to each extracted string. Declarative extraction does not implicitly coerce annotations.
• Text(..., strip=True) strips surrounding whitespace before transformation.
• Attr(..., absolute=True) resolves the attribute through response.url_join().

Plans are compiled from annotations once per Item class and then cached. A required-field or transform failure includes the item, field, selector, response URL, and root index. after_extract() is available for the irregular part of an otherwise declarative extraction:

class Article(Item):
    title: str = Text("h1")

    async def after_extract(self, response: HTMLResponse) -> None:
        self.title = self.title.strip()

Item.extract() intentionally does not replace Spider, callbacks, requests, middlewares, or pipelines. See examples/declarative_quotes_spider.py for a complete spider with pagination.

run_spider/crawl knobs:

• concurrency: number of concurrent HTTP requests; default 16; must be positive.
• max_pending_requests: queue bound to avoid unbounded memory use (defaults to concurrency * 10); if provided, must be positive. start_requests() always respects it, and so does a single producing callback; when several callbacks produce requests at once, they enqueue past it rather than parking workers or deadlocking.
• request_timeout: per-request timeout (seconds); defaults to 60, None disables it.
• keep_alive: reuse HTTP connections when supported by the underlying client (sends Connection: keep-alive).
• http_client: use a custom client instance such as OnionLinkClient(...) or ServoFetchClient(...) instead of the default wreq-backed client.
• dedup_key: optional Callable[[Request], str] used for request deduplication; defaults to lambda req: req.url.
• html_max_size_bytes: limit HTML parsed into AsyncDocument to avoid huge payloads.
• log_stats_interval: seconds between periodic stats logs; final stats are always emitted.
• request_middlewares / response_middlewares / item_pipelines: plug-ins run on every request/response/item.
• use run_spider_rsloop(...) instead of run_spider(...) to run under rsloop (requires pip install silkworm-rs[rsloop]).
• use run_spider_uvloop(...) instead of run_spider(...) to run under uvloop (requires pip install silkworm-rs[uvloop]).
• use run_spider_winloop(...) instead of run_spider(...) to run under winloop on Windows (requires pip install silkworm-rs[winloop]).

Built-in middlewares and pipelines

from silkworm.middlewares import (
    CloudflareCrawlMiddleware,
    CookiesMiddleware,
    DelayMiddleware,
    ProxyMiddleware,
    RequestResponseStreamMiddleware,
    RetryMiddleware,
    RobotsTxtDelayMiddleware,
    SkipNonHTMLMiddleware,
    UserAgentMiddleware,
)
from silkworm.pipelines import (
    CallbackPipeline,  # invoke a custom callback function on each item
    CSVPipeline,
    JsonLinesPipeline,
    MsgPackPipeline,  # requires: pip install silkworm-rs[msgpack]
    RssPipeline,
    SQLitePipeline,
    XMLPipeline,
    TaskiqPipeline,  # requires: pip install silkworm-rs[taskiq]
    ZenohPipeline,  # requires: pip install silkworm-rs[zenoh]
    PolarsPipeline,  # requires: pip install silkworm-rs[polars]
    ExcelPipeline,  # requires: pip install silkworm-rs[excel]
    YAMLPipeline,  # requires: pip install silkworm-rs[yaml]
    AvroPipeline,  # requires: pip install silkworm-rs[avro]
    ElasticsearchPipeline,  # requires: pip install silkworm-rs[elasticsearch]
    MongoDBPipeline,  # requires: pip install silkworm-rs[mongodb]
    MySQLPipeline,  # requires: pip install silkworm-rs[mysql]
    PostgreSQLPipeline,  # requires: pip install silkworm-rs[postgresql]
    S3JsonLinesPipeline,  # requires: pip install silkworm-rs[s3]
    VortexPipeline,  # requires: pip install silkworm-rs[vortex]
    WebhookPipeline,  # sends items to webhook endpoints using wreq
    GoogleSheetsPipeline,  # requires: pip install silkworm-rs[gsheets]
    SnowflakePipeline,  # requires: pip install silkworm-rs[snowflake]
    FTPPipeline,  # requires: pip install silkworm-rs[ftp]
    SFTPPipeline,  # requires: pip install silkworm-rs[sftp]
    CassandraPipeline,  # requires: pip install silkworm-rs[cassandra]
    CouchDBPipeline,  # requires: pip install silkworm-rs[couchdb]
    DynamoDBPipeline,  # requires: pip install silkworm-rs[dynamodb]
    DuckDBPipeline,  # requires: pip install silkworm-rs[duckdb]
)

run_spider(
    QuotesSpider,
    request_middlewares=[
        UserAgentMiddleware(),  # rotate/custom user agent
        DelayMiddleware(min_delay=0.3, max_delay=1.2),  # polite throttling
        # Read Crawl-delay/Request-rate from robots.txt and serialize same-origin requests
        # RobotsTxtDelayMiddleware("https://quotes.toscrape.com", user_agent="silkworm"),
        # ProxyMiddleware with round-robin selection (default)
        # ProxyMiddleware(proxies=["http://user:pass@proxy1:8080", "http://proxy2:8080"]),
        # ProxyMiddleware with random selection
        # ProxyMiddleware(proxies=["http://proxy1:8080", "http://proxy2:8080"], random_selection=True),
        # ProxyMiddleware from file with random selection
        # ProxyMiddleware(proxy_file="proxies.txt", random_selection=True),
    ],
    response_middlewares=[
        RetryMiddleware(max_times=3, sleep_http_codes=[403, 429]),  # backoff + retry
        SkipNonHTMLMiddleware(),  # drop callbacks for images/APIs/etc
    ],
    item_pipelines=[
        JsonLinesPipeline("data/quotes.jl"),
        SQLitePipeline("data/quotes.db", table="quotes"),
        XMLPipeline("data/quotes.xml", root_element="quotes", item_element="quote"),
        CSVPipeline("data/quotes.csv", fieldnames=["author", "text", "tags"]),
        MsgPackPipeline("data/quotes.msgpack"),
    ],
)

• DelayMiddleware strategies: delay=1.0 (fixed), min_delay/max_delay (random), or delay_func (custom).
• RobotsTxtDelayMiddleware("https://example.com", user_agent="silkworm") downloads https://example.com/robots.txt, applies Crawl-delay or Request-rate, and serializes matching-origin requests so concurrency cannot bypass the configured spacing. Use fallback_delay=... to keep a conservative delay when robots.txt cannot be fetched.
• ProxyMiddleware supports three modes:
  • Round-robin (default): ProxyMiddleware(proxies=["http://proxy1:8080", "http://proxy2:8080"]) cycles through proxies in order.
  • Random selection: ProxyMiddleware(proxies=["http://proxy1:8080", "http://proxy2:8080"], random_selection=True) randomly selects a proxy for each request.
  • From file: ProxyMiddleware(proxy_file="proxies.txt") loads proxies from a file (one proxy per line, blank lines ignored). Combine with random_selection=True for random selection from the file.
• CookiesMiddleware stores Set-Cookie response headers, applies matching Cookie request headers, supports named jars via request.meta["cookiejar"], per-request cookies via request.meta["cookies"], opt-out via request.meta["dont_merge_cookies"], and Netscape/Mozilla cookie file save(...)/load(...). Use the same instance in request_middlewares and response_middlewares.
• RetryMiddleware backs off with asyncio.sleep; any status in sleep_http_codes is retried even if not in retry_http_codes.
• SkipNonHTMLMiddleware checks Content-Type and optionally sniffs the body (sniff_bytes) to avoid running HTML callbacks on binary/API responses.
• CloudflareCrawlMiddleware is opt-in per request via request.meta["cloudflare_crawl"]; it submits a Cloudflare Browser Rendering crawl job, polls until completion, and hands your callback a synthetic JSON Response with the final API payload.
• RequestResponseStreamMiddleware streams paired request/response telemetry events to a collector endpoint; use the same instance in both middleware lists.
• JsonLinesPipeline writes items to a local JSON Lines file and, when opendal is installed, appends asynchronously via the filesystem backend (use_opendal=False to stick to a regular file handle).
• CSVPipeline flattens nested dicts (e.g., {"user": {"name": "Alice"}} -> user_name) and joins lists with commas; XMLPipeline preserves nesting.
• RssPipeline writes buffered RSS 2.0 feeds from items with configurable title/link/description fields.
• MsgPackPipeline writes items in binary MessagePack format using ormsgpack for fast and compact serialization (requires pip install silkworm-rs[msgpack]).
• TaskiqPipeline sends items to a Taskiq queue for distributed processing (requires pip install silkworm-rs[taskiq]).
• ZenohPipeline publishes JSON items to a static or dynamically resolved Zenoh key expression (requires pip install silkworm-rs[zenoh]).
• PolarsPipeline writes items to a Parquet file using Polars for efficient columnar storage (requires pip install silkworm-rs[polars]).
• ExcelPipeline writes items to an Excel .xlsx file (requires pip install silkworm-rs[excel]).
• YAMLPipeline writes items to a YAML file (requires pip install silkworm-rs[yaml]).
• AvroPipeline writes items to an Avro file with optional schema (requires pip install silkworm-rs[avro]).
• ElasticsearchPipeline sends items to an Elasticsearch index (requires pip install silkworm-rs[elasticsearch]).
• MongoDBPipeline sends items to a MongoDB collection (requires pip install silkworm-rs[mongodb]).
• MySQLPipeline sends items to a MySQL database table as JSON (requires pip install silkworm-rs[mysql]).
• PostgreSQLPipeline sends items to a PostgreSQL database table as JSONB (requires pip install silkworm-rs[postgresql]).
• S3JsonLinesPipeline writes items to AWS S3 in JSON Lines format using async OpenDAL (requires pip install silkworm-rs[s3]).
• VortexPipeline writes items to a Vortex file for high-performance columnar storage with 100x faster random access and 10-20x faster scans compared to Parquet (requires pip install silkworm-rs[vortex]).
• WebhookPipeline sends items to webhook endpoints via HTTP POST/PUT using wreq (same HTTP client as the spider) with support for batching and custom headers.
• GoogleSheetsPipeline appends items to Google Sheets with automatic flattening of nested data structures (requires pip install silkworm-rs[gsheets] and service account credentials).
• SnowflakePipeline sends items to Snowflake data warehouse tables as JSON (requires pip install silkworm-rs[snowflake]).
• FTPPipeline writes items to an FTP server in JSON Lines format (requires pip install silkworm-rs[ftp]).
• SFTPPipeline writes items to an SFTP server in JSON Lines format with support for password or key-based authentication (requires pip install silkworm-rs[sftp]).
• CassandraPipeline sends items to Apache Cassandra database tables (requires pip install silkworm-rs[cassandra]).
• CouchDBPipeline sends items to CouchDB databases as documents (requires pip install silkworm-rs[couchdb]).
• DynamoDBPipeline sends items to AWS DynamoDB tables with automatic table creation (requires pip install silkworm-rs[dynamodb]).
• DuckDBPipeline sends items to a DuckDB database table as JSON (requires pip install silkworm-rs[duckdb]).
• CallbackPipeline invokes a custom callback function (sync or async) on each item, enabling inline processing logic without creating a full pipeline class. See example below.

Using CallbackPipeline for custom processing

Process items with custom callback functions without creating a full pipeline class:

from silkworm.pipelines import CallbackPipeline

# Sync callback
def print_item(item, spider):
    print(f"[{spider.name}] {item}")
    return item

# Async callback
async def validate_item(item, spider):
    # Could do async operations like database checks
    if len(item.get("text", "")) < 10:
        print(f"Warning: Short text in item")
    return item

# Modifying callback
def enrich_item(item, spider):
    item["spider_name"] = spider.name
    item["processed"] = True
    return item

run_spider(
    QuotesSpider,
    item_pipelines=[
        CallbackPipeline(callback=print_item),
        CallbackPipeline(callback=validate_item),
        CallbackPipeline(callback=enrich_item),
    ],
)

Callbacks receive (item, spider) and should return the processed item (or None to return the original item unchanged).

Streaming items to a queue with TaskiqPipeline

Stream scraped items to a Taskiq queue for distributed processing:

from taskiq import InMemoryBroker
from silkworm.pipelines import TaskiqPipeline

broker = InMemoryBroker()

@broker.task
async def process_item(item):
    # Your item processing logic here
    print(f"Processing: {item}")
    # Save to database, send to another service, etc.

pipeline = TaskiqPipeline(broker, task=process_item)
run_spider(MySpider, item_pipelines=[pipeline])

This enables distributed processing, retries, rate limiting, and other Taskiq features. See examples/taskiq_quotes_spider.py for a complete example.

Publishing items with ZenohPipeline

Publish every item immediately to a Zenoh key expression:

from silkworm.pipelines import ZenohPipeline

pipeline = ZenohPipeline("scraping/quotes")
run_spider(MySpider, item_pipelines=[pipeline])

Use a sync or async resolver when items need separate keys:

async def item_key(item, spider):
    return f"scraping/{spider.name}/{item['category']}"

pipeline = ZenohPipeline(item_key)

The pipeline opens and closes its own session by default. Pass a configured zenoh.Config instance to customize that session, or session=existing_session to reuse a caller-owned session. Injected sessions remain open when the pipeline closes. Publisher options include encoding, congestion_control, priority, express, reliability, and allowed_destination. Dynamic publishers are cached by key until pipeline shutdown.

Handling non-HTML responses

Keep crawls cheap when URLs mix HTML and binaries/APIs:

response_middlewares=[SkipNonHTMLMiddleware(sniff_bytes=1024)]
# Tighten HTML parsing size (bytes) to avoid loading huge bodies into scraper-rs
run_spider(MySpider, html_max_size_bytes=1_000_000)

Performance optimization with rsloop

For improved async performance, enable rsloop as a drop-in replacement for asyncio's event loop:

pip install silkworm-rs[rsloop]
# or with uv:
uv pip install silkworm-rs[rsloop]

Then call run_spider_rsloop (same signature as run_spider):

from silkworm import run_spider_rsloop

run_spider_rsloop(
    QuotesSpider,
    concurrency=32,
)

Performance optimization with uvloop

For improved async performance, enable uvloop (a fast, drop-in replacement for asyncio's event loop):

pip install silkworm-rs[uvloop]
# or with uv:
uv pip install silkworm-rs[uvloop]

Then call run_spider_uvloop (same signature as run_spider):

from silkworm import run_spider_uvloop

run_spider_uvloop(
    QuotesSpider,
    concurrency=32,
)

uvloop can provide 2-4x performance improvement for I/O-bound workloads.

Performance optimization with winloop (Windows)

For Windows users who want improved async performance, enable winloop (a Windows-compatible alternative to uvloop):

pip install silkworm-rs[winloop]
# or with uv:
uv pip install silkworm-rs[winloop]

Then call run_spider_winloop (same signature as run_spider):

from silkworm import run_spider_winloop

run_spider_winloop(
    QuotesSpider,
    concurrency=32,
)

winloop provides significant performance improvements on Windows, similar to what uvloop offers on Unix-like systems.

Running spiders with trio

If you prefer trio over asyncio, you can use run_spider_trio instead of run_spider:

pip install silkworm-rs[trio]
# or with uv:
uv pip install silkworm-rs[trio]

Then use run_spider_trio:

from silkworm import run_spider_trio

run_spider_trio(
    QuotesSpider,
    concurrency=16,
    request_timeout=10,
)

This runs your spider using trio as the async backend via trio-asyncio compatibility layer. The Trio runner currently requires Python 3.13 because trio-asyncio 0.16 is not compatible with Python 3.14 or newer.

JavaScript rendering with Servo

For pages that need JavaScript execution but do not require driving an external browser process, install the optional Servo renderer and pass ServoFetchClient as the spider HTTP client.

Install a wheel from this page: https://github.com/RustedBytes/servofetch-py/releases

from silkworm import HTMLResponse, Response, ServoFetchClient, Spider, run_spider


class RenderedSpider(Spider):
    name = "rendered"
    start_urls = ("https://example.com/",)

    async def parse(self, response: Response) -> None:
        if isinstance(response, HTMLResponse):
            title = await response.select_first("title")
            await self.emit({"title": title.text if title else ""})


run_spider(RenderedSpider, http_client=ServoFetchClient(settle_ms=500))

Per-request render options live in Request.meta: servo_javascript, servo_settle_ms, servo_user_agent, servo_screenshot, and servo_full_page. Request.timeout overrides the client timeout for that request.

ServoFetchClient embeds Servo through servofetch; the existing CDP client connects to an external Lightpanda/Chrome-compatible browser over WebSocket. Use the default wreq client when pages do not need client-side rendering.

JavaScript rendering with Lightpanda (CDP)

For pages that require JavaScript execution, you can use Lightpanda (or any CDP-compatible browser) instead of the standard HTTP client. This uses the Chrome DevTools Protocol (CDP) to control a browser.

Installation

pip install silkworm-rs[cdp]
# or with uv:
uv pip install silkworm-rs[cdp]

Starting Lightpanda

lightpanda --remote-debugging-port=9222

Or use Chrome/Chromium:

chromium --remote-debugging-port=9222 --headless

The same ws://127.0.0.1:9222 endpoint works for both. Chrome only accepts its per-session ws://.../devtools/browser/<id> URL, so when a bare host:port endpoint is refused, CDPClient looks that URL up at http://host:port/json/version and connects to it (keeping your host and port). You can also pass the full URL yourself.

Using CDP in your spider

There are two ways to use CDP: the convenience API or custom spider integration.

Convenience API (simple one-off fetches)

import asyncio
from silkworm import fetch_html_cdp

async def main():
    # Fetch HTML with JavaScript rendering
    text, doc = await fetch_html_cdp(
        "https://example.com",
        ws_endpoint="ws://127.0.0.1:9222",
        timeout=30.0
    )
    
    # Extract data from rendered page
    title = await doc.select_first("title")
    print(title.text if title else "No title")

asyncio.run(main())

Full Spider Integration

from silkworm import HTMLResponse, Request, Response, Spider
from silkworm.cdp import CDPClient

class LightpandaSpider(Spider):
    name = "lightpanda"
    start_urls = ("https://example.com/",)

    def __init__(self, **kwargs):
        super().__init__(**kwargs)
        self._cdp_client = None

    async def start_requests(self) -> None:
        # Connect to CDP endpoint
        self._cdp_client = CDPClient(
            ws_endpoint="ws://127.0.0.1:9222",
            timeout=30.0
        )
        await self._cdp_client.connect()
        
        for url in self.start_urls:
            await self.follow(url, callback=self.parse)

    async def parse(self, response: Response) -> None:
        if not isinstance(response, HTMLResponse):
            return
        
        # Extract links from JavaScript-rendered page
        for link in await response.select("a"):
            href = link.attr("href")
            if href:
                await self.emit({"url": href})

    async def close(self):
        if self._cdp_client:
            await self._cdp_client.close()

See examples/lightpanda_simple.py and examples/lightpanda_spider.py for complete working examples.

Note: CDP support is experimental. For production use, consider using dedicated browser automation tools or the standard HTTP client when JavaScript rendering is not required.

Onion services with OnionLink

For Tor v3 .onion sites, install the optional OnionLink extra and pass OnionLinkClient as the spider HTTP client:

pip install "silkworm-rs[onionlink]"

from silkworm import HTMLResponse, OnionLinkClient, Response, Spider, run_spider


class OnionSpider(Spider):
    name = "onion"
    start_urls = ("http://exampleexampleexampleexampleexampleexampleexampleexampleexampleexample.onion/",)

    async def parse(self, response: Response) -> None:
        if isinstance(response, HTMLResponse):
            title = await response.select_first("title")
            await self.emit({"title": title.text if title else ""})


run_spider(
    OnionSpider,
    http_client=OnionLinkClient(concurrency=4, timeout=30),
)

OnionLinkClient supports Silkworm Request headers, params, body/data, JSON payloads, redirects, HTML detection, and request.meta["redirect_times"]. Override OnionLink's response byte cap per request with request.meta["onionlink_response_limit"].

Logging and crawl statistics

• Structured logs via the standard library; set SILKWORM_LOG_LEVEL=DEBUG for verbose request/response/middleware output.
• Periodic statistics with log_stats_interval; final stats always include the close reason, elapsed time, queue size, requests/sec, seen URLs, items scraped and dropped, errors, retries, filtered requests, memory MB, and per-status/domain/error breakdowns. The same data is returned as a CrawlResult and can be served as Prometheus metrics (metrics_port).

Limitations

• By default, HTTP fetches are wreq-based without JavaScript execution; pages requiring client-side rendering can use the optional CDP integration (see "JavaScript rendering with Lightpanda" section) or external browser automation tools. Tor v3 .onion sites can use the optional OnionLink integration.
• Request deduplication uses the request fingerprint (method, canonical URL with params, and body); headers and meta are ignored, so requests that differ only in headers are dropped unless you set dont_filter=True or pass a custom dedup_key.
• Redirects are followed inside the HTTP client, so allowed_domains filters requests before they are sent but a redirect can still land on another host.
• HTML parsing auto-detects encoding (BOM, HTTP headers/meta, charset detection fallback) but still enforces a html_max_size_bytes/doc_max_size_bytes cap (default 5 MB) in scraper-rs selectors, so very large pages may need a higher limit or preprocessing.
• Several pipelines buffer all items in memory until close (PolarsPipeline, ExcelPipeline, YAMLPipeline, AvroPipeline, VortexPipeline, S3JsonLinesPipeline, FTPPipeline, SFTPPipeline), which can bloat RAM on long crawls; prefer streaming pipelines like JsonLines/CSV/SQLite for high-volume runs.
• Many destination pipelines rely on optional extras; CassandraPipeline is disabled on Windows because cassandra-driver depends on libev there.

Examples

• python examples/quotes_spider.py → data/quotes.jl
• python examples/quotes_spider_trio.py → data/quotes_trio.jl (demonstrates trio backend)
• python examples/quotes_spider_winloop.py → data/quotes_winloop.jl (demonstrates winloop backend for Windows)
• python examples/hackernews_spider.py --pages 5 → data/hackernews.jl
• python examples/lobsters_spider.py --pages 2 → data/lobsters.jl
• python examples/start_urls_from_file_spider.py --urls-file data/start_urls.txt --output data/start_urls_from_file.jl (reads one URL per line and schedules custom requests with await self.follow(...) in start_requests)
• python examples/url_titles_spider.py --urls-file data/url_titles.jl --output data/titles.jl (includes SkipNonHTMLMiddleware and stricter HTML size limits)
• python examples/exception_handling_spider.py → data/exception_handling.jl (demonstrates process_exception and request errback)
• python examples/cookie_reuse_spiders.py → data/cookie_reuse.jl and data/cookies.txt (captures cookies in one run, saves them, then loads them for a second run)
• SILKWORM_LOG_LEVEL=DEBUG python examples/logging_controls_demo.py --mode noisy then --mode quiet → demonstrates noisy pipeline/URL logging and the quieter EngineLogger + pipeline log_level=None setup
• python examples/export_formats_demo.py --pages 2 → JSONL, XML, and CSV outputs in data/
• python examples/taskiq_quotes_spider.py --pages 2 → demonstrates TaskiqPipeline for queue-based processing
• python examples/sitemap_spider.py --sitemap-url https://example.com/sitemap.xml --pages 50 → data/sitemap_meta.jl (extracts meta tags and Open Graph data from sitemap URLs)
• python examples/request_response_stream_spider.py --collector-url https://collector.example.com/events → streams request/response telemetry while writing quotes output
• CLOUDFLARE_ACCOUNT_ID=... CLOUDFLARE_API_TOKEN=... python examples/cloudflare_crawl_spider.py https://example.com --limit 10 → submits a Cloudflare Browser Rendering crawl job
• python examples/lightpanda_simple.py → demonstrates CDP/Lightpanda for JavaScript rendering (requires pip install silkworm-rs[cdp] and running Lightpanda)
• python examples/lightpanda_spider.py → full spider example using CDP/Lightpanda
• python examples/servo_spider.py → full spider example using ServoFetchClient and a servofetch wheel

Convenience API

For one-off fetches without a full spider:

Standard HTTP fetch

import asyncio
from silkworm import fetch_html

async def main():
    text, doc = await fetch_html("https://example.com")
    title = await doc.select_first("title")
    print(title.text if title else "No title")

asyncio.run(main())

CDP-based fetch (with JavaScript rendering)

import asyncio
from silkworm import fetch_html_cdp

async def main():
    # Requires Lightpanda/Chrome running with CDP enabled
    text, doc = await fetch_html_cdp("https://example.com")
    title = await doc.select_first("title")
    print(title.text if title else "No title")

asyncio.run(main())

Servo-based fetch

import asyncio
from silkworm import fetch_html_servo

async def main():
    # Requires a compatible servofetch wheel.
    text, doc = await fetch_html_servo("https://example.com", settle_ms=500)
    title = await doc.select_first("title")
    print(title.text if title else "No title")

asyncio.run(main())

HTML to Markdown

from silkworm import HTMLResponse, Response, Spider, html_to_markdown, stream_html_to_markdown

markdown = html_to_markdown("<h1>Hello</h1><p>World</p>", mode="minimal")
streamed = stream_html_to_markdown(["<h1>Hello</h1>", "<p>World</p>"])


class MarkdownSpider(Spider):
    name = "markdown"
    start_urls = ("https://example.com",)

    async def parse(self, response: Response) -> None:
        if not isinstance(response, HTMLResponse):
            return

        await self.emit(
            {
                "url": response.url,
                "markdown": await response.to_markdown(mode="full"),
            }
        )

Modes are full for rich conversion, minimal for the lean Fast DOM path, and mdream for the mdream-backed converter. to_markdown_result(...) and convert_html_to_markdown(...) return fast-h2m's structured result.

Contributing

Pull requests and issues are welcome. To set up a dev environment, install uv, create a Python 3.13 virtualenv, and sync dev dependencies:

uv venv --python python3.13
uv sync --group dev

Run the checks before opening a PR:

just fmt && just lint && just typecheck && just test

Acknowledgements

Silkworm is built on top of excellent open-source projects:

• wreq - HTTP client with browser impersonation capabilities
• onionlink - Tor v3 onion-service client
• servofetch - Bindings to the Servo browser
• scraper-rs - Fast HTML parsing library
• fast-h2m - Fast HTML-to-Markdown conversion
• rxml - XML parsing and writing

We are grateful to the maintainers and contributors of these projects for their work.

License

MIT License. See LICENSE for details.