Netweir

An ultra-fast, self-healing, undetectable web scraper. Rust does the work; you write Python.

That's the finished product, and it isn't finished yet. Here's what each of those words will mean when it is:

• Ultra-fast. Fetching, parsing, crawl scheduling and extraction all run in Rust, off Python's GIL. A spider that only declares what it wants never touches Python per page. Every speed claim is a benchmark in this repository that you can run.
• Self-healing. Selectors that find their element again after a site redesign and tell you they had to. Blocks spotted and recovered from by rotating, backing off or slowing down. Crawls that pick up where they stopped after a crash.
• Undetectable. Requests that are byte-for-byte what a real browser sends, from the TLS handshake to the order of the headers, and a real Chrome for the pages that need one.

Working today:

• Fetching that is indistinguishable from Chrome 155, Firefox 158 or Safari 27 on the wire, over HTTP/2 and HTTP/1.1, with cookies and redirects. Below: how that's proved, and the one known exception.
• Forms, logins and JSON APIs: POST and any other method, sent the way Chrome, Firefox or Safari sends a form or a page's own script, and forms read off a page the way a browser submits them. See docs/forms.md.
• The options requests and httpx users reach for, under the same names: params=, auth= (HTTP Basic, kept to the site you gave it to), allow_redirects=False to read a redirect, max_redirects, a timeout per call, and verify=False or a CA file for hosts with certificates of their own. In a crawl, a request takes them too, Scrapy's meta keys (download_timeout, dont_redirect, http_user...) work, and retry_http_codes, max_redirects and http_auth are settings. See docs/fetching.md.
• Sessions you control: a client's cookies to read, set and save to a storage state file Playwright reads too; a login done in Chrome carried on over HTTP, or the reverse; and several logged-in users in one crawl with Scrapy's cookies= and meta["cookiejar"], kept in the checkpoint. See docs/sessions.md.
• A parser that keeps pace with selectolax, the one to beat in Python, and pulls data out of the page faster.
• CSS queries with Scrapy's ::text and ::attr(), all of XPath 1.0, and Beautiful Soup's find_all family, each faster than the library you'd otherwise use for it.
• XML read as XML: RSS, Atom, sitemaps and SOAP, with namespaces, register_namespace() and remove_namespaces() as in Scrapy, and hostile XML (external entities, a "billion laughs") refused. Scrapy's XMLFeedSpider and CSVFeedSpider, with a node-by-node reader in Rust for feeds too big to hold. See docs/xml.md.
• Changing the page as Beautiful Soup does (decompose(), extract(), replace_with(), unwrap(), new_tag(), append(), prettify()), find_similar() to get every product card once you've found one, and JMESPath queries over JSON APIs and the JSON in a page's scripts. See docs/selecting.md.
• Structured data: page.metadata() reads a page's JSON-LD, microdata, Open Graph, Twitter cards and Dublin Core in one call, in Rust.
• Prices, dates and script data: parse_price gives a Decimal and the currency from £1,234.50 or 1.234,50 €; parse_date reads ISO, HTTP and numeric dates, month names in seven languages and 3 days ago; page.js_data("__INITIAL_STATE__") reads the JavaScript objects pages keep their data in. All three work as into= in an item. See docs/selecting.md.
• Item loaders, as Scrapy's: add_css(), add_xpath(), add_jmes(), MapCompose, TakeFirst and the rest, into dicts, dataclasses, attrs classes or Scrapy Items. Change the imports and your loaders carry over; they load a page's items in about three quarters of the time itemloaders and parsel take. See docs/loaders.md.
• Articles and Markdown: page.main_content() finds the story, post or description among the menus, sidebars, comments and ads (or tells you a listing has none), page.article() gives its title, byline, date, image and text, and page.markdown() writes the page or just its content as Markdown, tables and code blocks included. All in Rust, in a few milliseconds on a 1 MB page. See docs/selecting.md.
• Chunks for RAG: page.chunks() cuts the content's Markdown at its headings into pieces sized for an embedding model, keeps code blocks and tables whole where they fit, and gives each chunk its heading path and an id that stays the same from crawl to crawl. to_langchain() and to_llamaindex() hand them on, and netweir crawl docs --url URL -o docs.jsonl chunks a whole docs site. See docs/selecting.md.
• For AI agents: netweir mcp is an MCP server. Add it to Claude Code, Claude Desktop or Cursor and the agent can fetch pages as Chrome would and read them as Markdown, chunks, or products, jobs and articles as data. See docs/mcp.md.
• Auto-extraction: page.extract(netweir.auto.Product) reads a product, a category page's listing, a job posting, an article, a forum thread, a business's page, a property, a social post or a search engine's results from the structured data the page carries (JSON-LD, microdata, Open Graph), the state apps keep in their scripts (Next.js, Nuxt, Apollo), and its layout where those run out, in the common item schema Zyte's API uses, with where each field came from. When one field is wrong for a site, pin it with a selector and the rest stay automatic; it works in declarative spiders, and netweir extract URL tries it from the command line. netweir crawl ecommerce --url URL (or article, job) crawls a whole site with it, finding the categories and pages itself. See docs/extraction.md.
• Language-model extraction, if you want it: netweir.llm.extract(page, {"name": "the loaf's name", "weight": "its weight"}) has Claude read the fields you describe from a page no type covers, in exactly the shape you asked for. Your key, and nothing calls a model unless you ask. See docs/llm.md.
• Crawling: spiders, sitemaps, robots.txt, per-site throttling, and items written to JSON Lines, JSON, XML, CSV or Parquet, on disk or in cloud storage, in one file or in batches. Downloader and spider middlewares, most of Scrapy's signals, from_crawler, and pipelines with open_spider/close_spider work as Scrapy's do, so most Scrapy pipelines and middlewares carry over unchanged. allowed_domains keeps a crawl on the sites you meant; max_items, max_pages, max_errors, max_time and max_memory stop it when you've seen enough.
• Files and images an item links to, downloaded through the crawl itself (its robots.txt, delays, proxies and cache) and stored once each, on disk or in a bucket, in Scrapy's layout and with Scrapy's file_urls and image_urls fields. Images become JPEGs with thumbnails. See docs/crawling.md.
• netweir serve: spiders run as jobs you schedule, watch and cancel over HTTP, with their logs, items and stats kept, like scrapyd and speaking much of its API. See docs/serve.md.
• netweir shell URL, fetch and view: the page Netweir gets, at a Python prompt, on stdout or in your browser, for trying selectors before they go in a spider. See docs/shell.md.
• Scrapy's project commands, so your hands already know them: startproject, genspider (plain, rules, sitemap or feed spiders), list, crawl NAME, parse URL to see what a callback makes of one page, and check, which runs Scrapy-style contracts from your docstrings, custom ones included, and exits 1 when a redesign broke a spider. See docs/project.md.
• Validating items against a dataclass or a JSON Schema, and checks at the end of a crawl (item counts, error rate, field coverage) that make netweir crawl exit 1 when a crawl finished but didn't work.
• Watching a crawl while it runs: a progress line in the log every minute, live spider.stats with Scrapy's inc_value and friends, and an optional endpoint Prometheus can scrape. See docs/crawling.md.
• A download handler for Scrapy, so a Scrapy project can fetch as a browser does without moving off Scrapy. See docs/crawling.md.
• A response cache for the days you spend writing a spider: run it thirty times and the site serves its pages once. See docs/crawling.md.
• A crawl you run every day can skip the pages whose items it already has, as scrapy-deltafetch does, and still read every listing for what's new. The output keeps every run's items, a crash included. See docs/crawling.md.
• WARC and HAR files of everything a crawl fetched: the requests as they went out, header for header, and the responses, redirect hops and robots.txt included; logins are left out. Read them back with warcio, or open them in a browser's network panel. See docs/crawling.md.
• Declarative spiders, where you describe the item and the links and Rust does the rest without running any Python per page.
• Callbacks in worker processes, for spiders whose own Python is the slow part: workers=4 ran a heavy spider 3.1 times faster, with the same output.
• One crawl on many machines: processes share a request queue and a duplicate filter in Redis, as with scrapy-redis, and a process that dies has its requests picked up by the others. Feed a running crawl with netweir queue push. See docs/distributed.md.
• Self-healing: block pages recognised and recovered from, crawls that resume after a crash, and selectors that find their element again after a redesign.
• Proxy pools: a new proxy for every request, one per site, or one only after a block, with proxies that keep failing benched for a while. See docs/self-healing.md.
• A Chrome driver for pages that need JavaScript: click, type, hover, scroll an endless feed to its end, pick from dropdowns, tick boxes, upload files and reach into iframes like Playwright, finding elements by CSS, XPath, their text or their role, save a page as a PDF, replay a HAR file, record a trace, then read the page with Netweir's selectors, or read the JSON its own requests got back. A profile can be kept between runs. Requests can be blocked, changed or answered before they leave Chrome. An alert() never stalls a page, popups and new tabs come back as pages you can read, and downloads land in a folder you can save from. In a crawl, it can take over a request a site keeps blocking, and hand the cookies it earns back to the fast HTTP client. netweir install chrome fetches the Chrome it's tuned for.
• Steps: a request, a rule or a whole crawl can carry a list of things for Chrome to do before the callback reads the page, such as dismissing a banner, clicking "Load more", typing a search, scrolling a feed to its end, taking a screenshot or running your own function. The list is JSON, so it survives checkpoints and shared queues. Each step's outcome comes back to the callback, and the input is the same human mouse, touch and keys as the driver's. See docs/steps.md.
• A page's environment, with Playwright's names: viewport and screen, phones (device="Galaxy S23"), language, time zone, location and permissions, dark mode, extra headers, offline and init scripts, per page, per context, for a whole browser or for a crawl. Each is set where Chrome sets it for itself, so the page, its workers and frames from other sites all say the same thing, and the request headers agree. See docs/environment.md.

Coming next: Firefox in the driver, once there's a Firefox that doesn't announce it's automated. The designs are in docs/design/.

pip install netweir

There are wheels for Linux, macOS and Windows, on Python 3.11 and up, free-threaded 3.14 included, so there's nothing to compile. Pages that need a real browser want Chrome too: netweir install chrome fetches the version Netweir is tuned for.

import netweir

page = netweir.get("https://books.toscrape.com/")

page.css("h1::text").get()  # "All products"
page.css(".price_color::text").getall()  # ["£51.77", "£53.74", ...]
page.css("article h3 a::attr(href)").getall()  # ["/catalogue/...", ...]

for book in page.css("article.product_pod"):
    print(book.css("h3 a::attr(title)").get(), book.css(".price_color::text").get())

If you've used Scrapy, ::text, ::attr(), get() and getall() mean what you think they mean. Already have the HTML? netweir.parse(html) skips the request and gives you something you query the same way.

Whatever you scrape with now, your selectors should work here unchanged. XPath is all of XPath 1.0, plus the extras parsel users lean on:

page.xpath("//article//h3/a/@title").getall()
page.xpath("//p[has-class('price_color')]/text()").get()
page.xpath("//article//a[re:test(@href, '_\\d+/index\\.html$')]/@href").getall()
page.xpath("//li[@class=$cls]", cls="next").get()  # variables, as in parsel
page.css(".price_color::text").re(r"[\d.]+")  # ["51.77", "53.74", ...]

And if you think in Beautiful Soup, the same page answers that way too:

import re

page.find("ul", class_="pager").find("a")["href"]
page.find_all("a", title=True, limit=10)
page.find_all(string=re.compile("£"))
page.find("h3").find_parent("li").find_next_sibling("li")

Two things work slightly differently, on purpose. get() on an element gives its HTML the way Chrome's outerHTML writes it, so an attribute holding a quote comes out as " where lxml would switch to single quotes; the elements and their order are the same. And attribute values are plain strings: node["class"] is "star-rating Three", not Beautiful Soup's list. Filters still match one class out of several, as you'd expect.

Most shops, job boards and news sites describe their pages to search engines in structured data. Netweir reads it, falls back on the page's layout for whatever it leaves out, and gives you the item:

product = page.extract(netweir.auto.Product)
product["name"], product["price"], product["currency"]
product["metadata"]["sources"]["price"]  # "json-ld"

Field names and shapes are the common item schema's, so items from many sites, or from Zyte's API, share one table. When a site's data is wrong for one field, pin that field and keep the rest:

class Lumo(netweir.auto.Product):
    price = netweir.css("span.sale", track="price")

A pinned price element gives "9.50", not "$9.50 inc. VAT", and track= keeps the selector working through a redesign. Products, listings, job postings and articles are covered so far; see Auto-extraction.

A spider is a class with a parse method. It yields what it found and the links worth following:

import netweir


class Books(netweir.Spider):
    start_urls = ["https://books.toscrape.com/"]

    async def parse(self, page):
        for book in page.css("article.product_pod"):
            yield {
                "title": book.css("h3 a::attr(title)").get(),
                "price": book.css(".price_color::text").get(),
                "url": page.urljoin(book.css("h3 a::attr(href)").get()),
            }
        if next_page := page.css("li.next a::attr(href)").get():
            yield page.follow(next_page)

Run it from the shell, and the file extension picks the format:

netweir crawl books.py -o books.jsonl

That run fetched all 50 pages and wrote 1,000 books in 18 seconds, most of it spent waiting on purpose. Out of the box Netweir reads each site's robots.txt and stays out of what it disallows, skips pages whose owners reserve text and data mining rights (TDMRep), and spaces its requests to each site by how quickly the site answers, backing off when it gets a 429. A page linked three different ways, or with utm_ tags on the end, is fetched once.

All of that is a setting when you need it to be:

class Books(netweir.Spider):
    settings = netweir.Settings(concurrency=128, per_domain=4, throttle=False)
    pipelines = [drop_out_of_stock, netweir.export.parquet("books.parquet")]

A pipeline is any function, sync or async, that gets each item and returns it, changes it, or returns None to drop it. If you know Scrapy, the rest will look familiar: Request with callback, errback, meta and priority; page.follow_all(); -s concurrency=16 on the command line. A callback that raises is logged with the URL and counted, and the crawl carries on unless you set fail_fast=True.

Most spiders are the same three moves: follow the pagination, open each product, pull the same fields out of every one. You can say that instead of writing it:

class Book(netweir.Item):
    title = netweir.css("h1::text")
    price = netweir.css(".price_color::text", re=r"[\d.]+", into=float)
    upc = netweir.xpath("//th[.='UPC']/following-sibling::td/text()")
    stock = netweir.css(".availability::text", re=r"\d+", into=int)


class Books(netweir.Spider):
    start_urls = ["https://books.toscrape.com/"]
    rules = [
        netweir.Follow("li.next a"),
        netweir.Follow("article.product_pod h3 a", extract=Book),
    ]

Run against the real site, that crawled all 1,050 pages and wrote 1,000 books like this one:

{"title": "A Light in the Attic", "price": 51.77, "upc": "a897fe39b1053632", "stock": 22}

Every page in that crawl was parsed, searched and turned into an item in Rust, several pages at a time on separate threads. Python saw finished items and nothing else, which leaves your event loop and your GIL free for whatever your pipelines do. If a price won't turn into a float, you get None and one warning naming the field and the text it found, rather than a crash on page 600.

You don't have to choose. A rule can hand its pages to a callback, a callback can call Book.extract(page), and a spider can have both rules and its own parse. bench/rules.py crawls a local 1,050-page shop both ways; the declarative spider takes about two thirds of the time.

Three things go wrong on a long crawl, and Netweir handles each without being asked.

The site blocks you. Every response is checked against the block pages of Cloudflare, Akamai, DataDome, HUMAN, Kasada, Imperva and AWS WAF, so a challenge that comes back as a 403 isn't mistaken for a page. A blocked request is tried again with a new session (an empty cookie jar, and the next of your proxies if you gave several), and the site is slowed down. A server error is retried with backoff, a 429's Retry-After is honoured, and a site that keeps blocking is paused for a while instead of hammered. With browser="on_block", a request still blocked after all that gets one go in Chrome, which waits for the challenge to pass and then hands its cookies to the HTTP client, so the rest of the site doesn't need Chrome. If even that fails, your spider's on_block(request, page) hears about it. Outside a crawl, netweir.get raises netweir.Blocked, naming the vendor.

The crawl dies. Give it somewhere to keep its state and run it again after a crash, a reboot or a kill -9:

netweir crawl books.py -o books.jsonl -s checkpoint=crawls/books

It picks up where it stopped. Pages already done aren't fetched again, and no item ends up in books.jsonl twice: the tests kill a crawl at random moments until it finishes and check exactly that.

The site is redesigned. Name a selector and Netweir remembers what its element looked like:

price = page.css(".price_color::text", track="price")

When a new build renames price_color, wraps it in another div or swaps h1 for h2, the selector stops matching, and Netweir finds the element most like the one it remembers. price.relocated says it had to, and price.score says how sure it is. You get a warning, not an empty column. track= works on Item fields too. Add repair = netweir.repair.llm() to a spider, and a language model proposes a replacement selector for each one that broke. Netweir checks every proposal on the page and reports it to you, and never applies one itself.

Most sites don't block scrapers by reading their code. They block them by the first few hundred bytes of the connection: which TLS ciphers and extensions arrive, what the HTTP/2 settings frame says, which headers come first. A Python HTTP library gives itself away before it has asked for anything.

Netweir sends what Chrome 155 sends. Not something close to it: the same ClientHello, the same HTTP/2 settings and priorities, the same headers in the same order and the same capitalisation over HTTP/1.1. Cookies go where Chrome puts them, and redirects are followed hop by hop the way Chrome follows them, header order after a redirect included.

You don't have to take that on trust. Chrome was recorded doing four navigations against a local server: a page that sets a cookie, a revisit, a redirect to another site and one within the site. cargo test -p netweir-core has Netweir do the same four and fails if any request differs from Chrome's. The public checker at tls.peet.ws reports the same JA4 fingerprint for both: t13d1517h2_8daaf6152771_cb7bf5808d99.

Firefox 158 is there too: netweir.get(url, profile="firefox"), or profile="firefox" in a spider's settings. Firefox sends its TLS extensions in a fixed order, with no GREASE, and the same test checks that order against a recording of the real browser, as well as everything above. Haven't updated yet? The versions before stay, as chrome-154-macos and firefox-156-macos, and are tested the same way.

So is Safari 27, as profile="safari". It has two habits of its own, and Netweir copies both. Two cookies with the same path go out newest first. And once a site has answered over HTTP/1.1, Safari stops offering it HTTP/2 on new connections, so Netweir does the same.

One known difference: Chrome sends a priority header only over HTTP/2, and Netweir can't tell a server lacks HTTP/2 until it has answered once. So the first request to an https server that only speaks HTTP/1.1 carries that header; every request after it doesn't.

Images a crawl downloads with netweir.images go out as an <img> on the page asks for them, not as a page, and the same test holds Chrome's and Firefox's to recordings of each loading an image. Safari's are written from what Safari is known to send, and not yet checked against a recording.

A login goes where the browser puts it. Chrome 155 and Firefox 158 were recorded holding an HTTP login and sending it with a navigation, a form, a script's request and an image; Netweir's Authorization header lands in the same place in each, and the same test checks it. Safari's place isn't recorded yet, so with profile="safari" it goes after Safari's own headers. And verify=False or a CA file of your own only changes what Netweir accepts from the server, never what it sends: a test holds both handshakes to the browser's recording.

For many pages at once, use a client. It keeps connections and cookies between requests, and the requests run concurrently in Rust:

async with netweir.Client(proxy="http://user:pass@proxy:8080") as client:
    pages = await client.get_many(urls)

A request that fails before any response arrives raises netweir.FetchError, and its kind says why: "timeout", "connect", "tls" and so on. A 404 is still a page; check page.status.

Netweir parses with its own HTML parser, written in Rust from the HTML Standard. It passes the html5lib tokenizer tests and the web-platform-tests tree-construction tests that don't need JavaScript, and on 50 real pages it builds the same trees as lexbor, the C parser selectolax uses, about twice as fast. Then the whole query runs in Rust too, and Python gets a finished list of strings rather than an object for every element along the way. The GIL is released while it works, so threads parse in parallel.

bench/parse.py pulls every price, title and link out of a generated shop page:

Measured on an Apple M-series laptop with Python 3.14. Run it yourself: uv run --group bench python bench/parse.py.

XPath and find_all are a different story, because the libraries people use for them are much slower. bench/query.py parses the same shop page and asks six XPath questions, or six Beautiful Soup ones:

Look at lxml across the row: three times the page took ten times as long. Netweir builds a flat index of the document on the first query, so //x is a scan over a few arrays, and its time grows with the page and no faster.

Most spiders don't query the whole page, though. They loop over cards and ask each one for a few fields, and on an element that small, a query's cost is mostly overhead. bench/items.py does that for the shop page's 998 cards, four fields each:

You can write the query inline; each thread keeps the ones it has compiled. All three benchmarks run in CI, and the build fails if any library beats Netweir.

Pulling the readable part out of a page is the other job people time, usually right before it goes to a search index or a model. bench/content.py hands each tool a 1 MB article buried in navigation, a sidebar and a thousand comments, and times it from raw HTML to answer, parsing included:

That's twenty times faster on one M2 Max with Python 3.14, and CI fails the build if either one catches up. It measures speed, not judgement: trafilatura has years of tuning on news sites, so if what you need is the best possible text from a newspaper, test both on your own pages.

Parsing is one part of a crawl. bench/crawl.py times the whole thing: one local server plays 100 sites of 100 pages, each answer 50 ms late, and every crawler fetches all 10,000 pages and pulls a title and price from each, with the same limits (100 requests in flight, 8 per site):

With 100 requests in flight and 50 ms per answer, 2,000 pages a second is the most any crawler could do here, so Netweir is waiting on the server, not on itself. Scrapling is held back by its HTTP session's default of 10 connections, which it doesn't let you change; that is how it ships. Same laptop, Python 3.13; run it yourself with uv run python bench/crawl.py once the others are installed.

Pages you don't control can be built to be slow to parse. Pass a timeout and Netweir gives up instead of hanging:

netweir.parse(html, timeout=2.0)  # raises netweir.ParseTimeout

The timeout bounds time, not memory. A crawl reads at most 64 MiB of a response, counted after decompression, so a huge page or a small compressed one that unpacks to gigabytes fails instead of filling memory (max_response_size sets the limit; max_size= does the same for netweir.get). It doesn't stop a small page built to abuse the HTML5 spec's rules for misnested formatting tags, which can make any spec-compliant parser allocate gigabytes; nothing in Netweir caps the parser's memory yet.

Some pages are empty until their JavaScript runs. For those, Netweir drives Chrome, and you get the rendered page back as the same Node you'd get from a fetch:

async with netweir.browser() as browser:
    page = await browser.new_page()
    await page.goto("https://quotes.toscrape.com/js/")
    await page.click("li.next a")
    root = await page.parse()
    quotes = root.css("span.text::text").getall()

There's no sleep in that, and there doesn't need to be. A click waits until its button exists, is visible, has stopped moving, is enabled and isn't covered by a cookie banner, and if it never gets there, the error tells you which of those it was.

A driven Chrome normally gives itself away: navigator.webdriver is true, the user agent says HeadlessChrome, and the usual drivers switch on DevTools features that page scripts can notice. Netweir doesn't do any of that. A test serves a page that looks for the first two from script, and also fails if any of those DevTools features was switched on. docs/browser.md has the rest, including what it doesn't hide yet.

docs/ has a getting-started guide, a page for each part of Netweir and every setting with its default.

You need Rust, a C compiler, CMake and uv.

git clone --recurse-submodules https://github.com/netweir/netweir
cd netweir
uv sync --group dev
uv run maturin develop --uv
uv run pytest

uv installs the dependencies and maturin builds Netweir, so after changing any Rust, rerun uv run maturin develop --uv before testing.

If you set RUSTFLAGS yourself, it replaces the flags in .cargo/config.toml, so add --cfg pyo3_disable_reference_pool to it. Without that, threads on free-threaded Python barely scale, and tests/test_parse.py fails to tell you (docs/design/threads.md).

Issues, fixes and new browser profiles are all welcome. CONTRIBUTING.md covers building, testing and the one hard rule (write it yourself), and before your first pull request is merged you'll be asked to agree to the CLA, once. Questions go in Discussions, and ROADMAP.md says what's coming next.

AGPL-3.0. If that doesn't work for your company, a commercial licence is available: see netweir.com/licence or email legal@netweir.com. See NOTICE for the licences of what Netweir bundles.

The Netweir name and logo are trademarks of James Conway; the trademark policy says how you can use them. To sponsor Netweir, see GitHub Sponsors or email sponsors@netweir.com. Anything else: hello@netweir.com.

Please use it the way you'd want your own site scraped. Crawls obey robots.txt and TDMRep unless you switch that off, and Netweir tells you when you have. A one-off netweir.get() checks neither; that's on you.