ID@agntn/browsersv0.4.0

One scrape call. Eight browsers behind it.

Steel scrapes in one HTTP call. Kernel wants a live session first. Browserbase hands you a CDP URL. A model will mix those up, so this is one scrape() over all of them, in TypeScript, in the terminal and in your agent, and each backend says flag by flag what it can do.

Providers: 8; 13 flags each
Agent tools: 12; the same on MCP, Pi and OMP
Keys to start: 0; Playwright runs here

Get started Star on GitHub

Install$ pnpm add @agntn/browsers

Callbrowsers_scrape("https://example.com", "steel")recorded 2026-09-29 · 01 / 8

Provider / steel

Steel

One POST to /v1/scrape for a page, a session for everything else.

Route

createdone
sessionskipped
scrapedone
releaseskipped

Answer

<html lang="en"><head><meta charset="utf-8"><link rel="icon" href="data:,"><meta name="viewport" content="width=device-width,initial-scale=1"><title>Example Domain</title><style>html{color-scheme:light dark;background:light-dark(#eee,#222)}body{font:16px/1.6 system-ui,sans-serif;max-width:30em;min-height:100vh;margin:auto;padding:4.75em 2em 20vh;box-sizing:border-box;display:grid;place-content:center;text-align:center}</style></head><body><style>body{padding:2em 2em 20vh;place-items:center}svg{opacity:.55}p{grid-area:2/1}span{transition:opacity .4s}p+p span{opacity:0}@media(prefers-reduced-motion){*{transition:none}}</style><svg viewBox="0,0,20,20" width="44" height="44" fill="currentColor" aria-hidden="true"><path fill="none" stroke="currentColor" stroke-width="1.3" stroke-linejoin="round" d="M6 4H3v12q4 0 7 1.5-1-3.5-4-4.5V2q3.5 1 4 3v12.5q3-1.5 7-1.5V4q-4.5 0-7 1"></path><g transform="rotate(-6,13.6,8.3)"><path id="q" d="M11.5 6.6h1.8v1.8l-1 1.6-.6-.3.8-1.3h-1z"></path><use href="#q" x="2.4"></use></g></svg>This domain is for use in documentation examples without needing permission. This is not a service, avoid relying on it for testing and monitoring purposes.هذا النطاق مُخصص للاستخدام في أمثلة التوثيق دون الحاجة إلى إذن. هذه ليست خدمة، يُرجى تجنب الاعتماد عليها لأغراض الاختبار والمراقبة.该域名仅用于文档示例，无需获得许可。这并非一项服务，请勿将其用于测试和监控目的。L’usage de ce domaine est réservé à des exemples de documentation, sans autorisation préalable. Il ne s’agit pas d’un service ; son utilisation à des fins de test ou de surveillance est à éviter.Данный домен предназначен для использования в примерах документации без необходимости получения предварительного разрешения. Это не сервис; не рекомендуется его использование для тестирования и мониторинга.Este dominio está destinado al uso en ejemplos de documentación sin necesidad de permiso. Esto no es un servicio, evitar utilizarlo para realizar pruebas o monitoreos.<a href="https://iana.org/help/example-domains">Learn more</a><script src="/s.js"></script>
</body></html>

Came back: raw HTML
Size: 12,348 characters
Time: 1,915 ms
Session: none, one call

operations 4 / 13

capabilities() is the list

Six of the eight scrape in one call. Kernel only scrapes inside a session, which the tools and the CLI open and release for you. Anchor doesn't scrape at all, and says so instead of handing back an empty page. Everything past that is a flag per operation, read off a live instance, so a README can't drift from it. This grid is those flags.

Import(await create("cloudflare")).capabilities()

Callprovider.capabilities()8 providers · 13 operations

Operations per provider: one call, inside a session, or not supported
Provider	scrape	screenshot	element	pdf	crawl	links	a11y tree	extract	search	sessions	cdp	navigate	evaluate
steel	one call	inside a session	not supported	not supported	not supported	not supported	not supported	not supported	not supported	one call	one call	not supported	not supported
browserbase	one call	not supported	not supported	not supported	not supported	not supported	not supported	not supported	not supported	one call	one call	not supported	not supported
kernel	inside a session	inside a session	not supported	not supported	not supported	not supported	not supported	not supported	not supported	one call	one call	inside a session	inside a session
browserless	one call	one call	one call	one call	not supported	not supported	not supported	not supported	not supported	one call	one call	inside a session	inside a session
hyperbrowser	one call	one call	not supported	not supported	one call	not supported	not supported	one call	one call	one call	one call	not supported	not supported
anchor	not supported	inside a session	not supported	not supported	not supported	not supported	not supported	not supported	not supported	one call	one call	not supported	not supported
cloudflare	one call	one call	one call	one call	one call	one call	one call	one call	not supported	one call	one call	not supported	not supported
playwright	one call	inside a session	inside a session	one call	one call	one call	not supported	not supported	not supported	one call	not supported	inside a session	inside a session

Twelve tools, one executor

browsers mcp, the Pi extension and the OMP extension call the same executors, so a fix lands once and a model reads the same text on every host. The console asks browsers_capabilities about the provider in the panel above. The dialog holds exactly what a model would get.

A scrape stops at 20,000 characters unless you pass maxChars, and says how much it cut
Links come 500 at a time, with the offset to ask for next
A slow crawl hands back its job ID instead of a timeout

Read MCP, Pi and OMP

Callbrowsers_capabilities("steel")8 providers

Provider / steel

Steel

The flags come off a live instance, not a table in a README. What the model reads is what the next call will run into.

supports: 4 of 13 operations
scrape: one call
needs: STEEL_API_KEY

Bring your own backend

The built-ins are eight classes behind one interface, and yours is the same shape in one file. This one puts Jina Reader behind scrape(): one GET, markdown back, and an honest no for everything it can't do.

A class that implements BrowserProvider and a factory that builds it
notSupportedViaRest turns every missing operation into a typed refusal
register(name, defaultURL, factory) and create(name) finds it

Read Custom providers

Filereader.tsfolded · copy is whole

import { create, notSupportedViaRest, register } from "@agntn/browsers";class Reader implements BrowserProvider {  /* baseURL, name() and capabilities(): scrape only */  async scrape(url) { /* GET r.jina.ai/<url>, markdown back */ }  /* createSession … evaluate: notSupportedViaRest("reader", …) */}register("reader", "https://r.jina.ai", (config) => new Reader(config));const page = await (await create("reader")).scrape("https://example.com");

Startpnpm add @agntn/browsersNode.js 26 or newer

Start with the one that needs no key

One install gives you the library, the browsers CLI and the MCP server. Playwright runs on your machine, so the first scrape costs nothing. Later, set a vendor key and swap the name for resolveProvider(), and the call goes to the cloud.

PinPre-1.0, so pin exact versions.
ChromiumPlaywright takes the Chrome already on your PATH, or the one playwright-core installs.
BillsEvery vendor but Playwright meters its own account. The key is yours.

Read the guide Compare providers

First call

$ pnpm add @agntn/browsersimport { create } from "@agntn/browsers";const provider = await create("playwright");provider.capabilities().statelessScrape;  // trueconst page = await provider.scrape("https://example.com");page.text;  // "This domain is for use in documenta…"

One scrape call. Eight browsers behind it.

Steel

Browserbase

Kernel

Browserless

Hyperbrowser

Anchor

Cloudflare

Playwright

capabilities() is the list

Twelve tools, one executor

Steel

Bring your own backend

Start with the one that needs no key