Playwright automates a browser you script. Browsewright is the browser that scripts itself.

You don't write selectors. You don't maintain scrapers that break every time a site ships a redesign. You give it intent — "find the pricing", "enrich this lead", "fill out this form" — and an LLM drives a real browser to get it done.

pip install browsewright
bw "https://stripe.com" "what does this company do and who is it for"

============================================================
RESULT  [api]  412 tokens  3.1s
------------------------------------------------------------
Stripe is financial infrastructure for the internet. It provides
payment processing, billing, and treasury APIs for businesses from
startups to enterprises like Amazon and Shopify...
============================================================

🤯 It doesn't just read the web. It does things on the web.

Most "AI scrapers" hand you text. Browsewright acts. Point it at a real government records form with no API, give it a profile, and walk away:

bw-tasks form \
  "https://registers.maryland.gov/RowNetWeb/Estates/frmEstateSearch2.aspx" \
  --profile examples/sample_profile.json

It read the field labels, mapped your profile onto the form with an LLM, picked valid dropdown options, submitted it, and came back with:

Page 1 of 815 results — real names and dates, extracted as JSON.

No selectors. No XPath. No API. The form has none — it's a 20-year-old ASP.NET page that's invisible to every HTTP scraper. Browsewright drives it like a human.

💸 And it's almost free

Benchmark — 50 real, diverse websites in one run: 50 / 50 extracted successfully · $0.047 total · ~1,200 tokens & ~20s median per site. 28% were answered by the free API/archive shortcut with no browser at all. (Reproduce it: python examples/batch_test.py.)

It tries the cheapest path first — open APIs, RSS, public archives — and only spins up Chrome when a page actually needs it. You pay pennies for the easy 80% and a real browser for the hard 20%.

How it stacks up

Comparisons reflect typical default usage; all four are good tools. Browsewright's bet is intent in → action + structured data out, run locally for pennies.

Install

pip install browsewright          # core
pip install "browsewright[mcp]"   # + MCP server (Claude Desktop / Code / any client)

Or from source:

git clone https://github.com/krishnashakula/browsewright && cd browsewright
python -m venv .venv && . .venv/bin/activate    # Windows: .venv\Scripts\activate
pip install -e .

Add your Anthropic API key:

cp .env.example .env
# edit .env and paste your key from https://console.anthropic.com/settings/keys

The first browser run launches Chrome via nodriver (Chrome must be installed).

bw "not recognized" after install? pip put the scripts in a folder that isn't on your PATH (common on Windows). Use the module form, which always works: python -m browsewright "<url>" "<goal>" · python -m browsewright.tasks_cli enrich "<url>"

Use it

CLI

bw "https://news.ycombinator.com" "the top story right now"
bw "https://example.com" "find the pricing" --json
bw "https://example.com" "debug this" --no-headless --verbose

Python

import asyncio
from browsewright import search

res = asyncio.run(search("https://stripe.com", "what does this company do"))
print(res.answer)         # synthesized answer
print(res.stage)          # "api" | "browser" | "common_crawl" | "blocked" | "error"
print(res.tokens_total, res.elapsed_s)

As an MCP tool (Claude Desktop / Claude Code / any MCP client)

{ "mcpServers": { "browsewright": { "command": "bw-mcp" } } }

Your LLM now has a read_page(url, goal) tool.

The 5 built-in tasks — bw-tasks

One pipeline — fetch → structured extract (JSON) → diff/aggregate → action — exposed as five business workflows. Each is a CLI subcommand and a library function.

Common flags: --json, --out FILE, --slack <webhook>, --no-headless, --aggressive.

Real enrich output (trimmed):

{
  "company_name": "Tavily",
  "industry": "AI/SaaS - Developer Tools",
  "tech_stack_or_integrations": ["OpenAI", "Anthropic", "Groq", "Databricks"],
  "recent_news_or_signals": ["Raised $25M Series A", "Databricks MCP partnership"],
  "icp_fit_score_1_to_10": 7,
  "personalized_cold_email_first_line": "I noticed Tavily just partnered with Databricks on the MCP Marketplace—looks like you're doubling down on enterprise adoption after your $25M Series A."
}

Build your own task with the core primitive

Every task is a thin wrapper over extract_structured(url, schema). Define any schema, get JSON back:

import asyncio
from browsewright import extract_structured

schema = {"headline": "string",
          "open_roles": [{"title": "string", "team": "string", "location": "string"}]}
data = asyncio.run(extract_structured(
    "https://example.com/careers", schema,
    instruction="Extract the page headline and every open job posting."))
print(data["open_roles"])

Scheduling

Tasks are single-shot; snapshot/diff state persists between runs, so change detection works across invocations. Run on cron, n8n/Make/Zapier, or /loop:

# every 6h, alert on competitor pricing changes
0 */6 * * * bw-tasks watch "https://competitor.com/pricing" --slack https://hooks.slack.com/services/XXX

How it works

search(url, goal)
   │
   ├─ Polite gate ........ robots.txt check + per-host rate limit
   │
   ├─ Pre-flight pipeline (cheapest path first)
   │     1. Common Crawl ... public archive            (opt-in)
   │     2. Open API ....... RSS / wp-json / *.json     (no browser, ~1.5k tokens)
   │     3. Origin IP ...... CDN bypass                 (skipped in polite mode)
   │     4. Classifier ..... detect Cloudflare/Akamai/DataDome/…
   │
   └─ Browser session (only if no shortcut hit)
         • real headless Chrome via nodriver (native TLS fingerprint)
         • human motor layer — Bézier mouse, typing cadence, scroll pacing
         • LLM decides actions only at junctions (~1 call/page)
         • blind-scene shortcut: extract directly when the DOM scan is blocked
         • visual recovery: a vision call clears interstitials/challenges

Polite by default

Polite mode is the default and what you should ship. It checks robots.txt, rate-limits per host, and does not bypass CDN bot protection. --aggressive (polite=False) enables origin-IP discovery and ignores robots — use it only on targets you own or are authorized to test.

⚠️ You are responsible for complying with each site's Terms of Service, applicable law (CFAA and equivalents), and data-protection rules (GDPR/CCPA). Browsewright is for authorized research, your own properties, and sites whose terms permit automated access. The authors accept no liability for misuse.

⭐ Star it / contribute

If Browsewright saved you a scraper, drop a star — it's the whole reason this is open source. Issues and PRs welcome: pre-flight vendors, new tasks, more sites in the benchmark.

MIT licensed. Built on nodriver

• Anthropic Claude.