Browser Automation for Agents
Let an AI agent use websites like a person — navigate, click, fill forms and read pages — when there is no API.
Trends
Claude gained the most ground in package downloads: 4.4× against the median npm package over a year.
Public counts each month: pull requests on GitHub, package downloads, stars, Product Hunt followers. They show direction, not market share.
- Oct 19 · Pricing Claude changed its llm api pricing: Claude Haiku 4.5 plan dropped; new plan Claude Haiku 5.5. At the calculator's typical usage it now costs $10 a month instead of $100.
- Oct 7 · Action needed Claude API managed agents with limited networking now apply allowed_hosts to web_search and web_fetch, so unlisted hosts are blocked; add every host those tools need to allowed_hosts.
The real choice
Browser infrastructure or agent logic. Browser infrastructure runs the headless browsers for you, with sessions, proxies and recordings, and your code decides what to do. Agent libraries supply the model-driven logic that decides where to click, and can run on your own browser or a hosted one. Underneath both sits a plain automation library, where you script every step yourself.
Pick by situation
Tap the ones that are you — the tools that fit light up below.- The agent must handle pages it has never seen, given only a goal
Browser Use - You know most of the steps and want the model only for the parts that vary
Stagehand - The flow is fixed and can be scripted without a model at all
Playwright
Puppeteer - You need many parallel browsers, or sites block your server's IP
Browserbase
Steel
Compare the tools
Grouped by the side of the choice they answer, not ranked. Open a row for when to use it, the trade-off and what makers say.
Browser Uselibrary · open sourceGiving a model a browser and a goal in a few lines of Python, and letting it work out the clicks.

12+21 open source1.8M/wk
Use it whenYou want the agent to handle pages it hasn't seen before.
Trade-offEvery step costs model calls and time, and results vary run to run.
Used by Gedd.it, AgentMail, CirtusAI, ClawEase, CopyCat and 7 more · in 21 open-source projects.
Browser Use was the catalyst for AI Browser Agents. Some of our most critical functionality couldn't have been possible without them.
Stagehandframework · open sourceMixing scripted browser steps with natural-language act, extract and agent calls, in TypeScript or Python.10 open source1.5M/wk−9% vs npm
Use it whenMost of the flow is known, but some steps face pages that change or differ per site.
Trade-offWorks best on Browserbase, and each natural-language step still costs a model call.
No maker's product we track shows it for this yet · in 10 open-source projects.
BrowserbaseserviceRunning many headless browsers in the cloud with session recordings, proxies and captcha handling.


16+15 open source2.2M/wk+20% vs npm
Use it whenYour agents need to scale beyond one machine or get blocked from server IPs.
Trade-offHosted only, billed by browser time.
In ReplitUsed by Locus Founder, Spine, Walle, Brew, Caddy and 11 more · in 15 open-source projects.
Giving agents a real browser to work in changes what's possible. Locus Founder needs to operate in the real web, not just talk about it, and Browserbase is what makes that happen reliably.
Steelservice · open sourceBrowser infrastructure for agents that you can use as a cloud service or self-host from open source.4 open source34.3k/wk
Use it whenYou want hosted browsers now with the option to run them on your own servers later.
Trade-offSmaller ecosystem than larger browser-infrastructure providers.
No maker's product we track shows it for this yet · in 4 open-source projects.
Puppeteerlibrary · open sourceControlling headless Chrome from Node.js, for scraping, screenshots and PDF generation.112 open source20.4M/wk−37% vs npm
Use it whenYou work in JavaScript and only need Chrome or Firefox.
Trade-offJavaScript only, with fewer built-in waits and test tools than Playwright.
No maker's product we track shows it for this yet · in 112 open-source projects.
Playwrightframework · open sourceScripting Chromium, Firefox and WebKit with one API, in TypeScript, Python, Java or .NET.No evidence yet92.2M/wk+17% vs npm
Use it whenThe steps are known in advance, or you need the browser layer under an agent library.
Trade-offScripts break when page structure changes, and you run the browsers yourself.
No maker's product we track shows it for this yet.
Before you choose
Use an API if one exists; browser agents are slow, cost tokens on every step and break when pages change. Script the steps you know with plain Playwright and let the model handle only the parts that vary. Develop against a local browser, and move to hosted browsers when you need many parallel sessions or run from servers that sites block.
- Letting the model decide every click on a flow that never changes, paying tokens and latency for something a script does reliably.
- Running agents with a user's logged-in session and no limits on what they may submit, buy or delete.
- Ignoring a site's terms and bot protection, then building a feature that breaks the week it gets blocked.
Other options
Real choices most makers here won't need to weigh.
Decided alongside
What the 28 makers' products here chose for their other decisions.