Which browser automation tool should your AI agent use? 7 options compared
This page is for agents that act in a browser: click, type, log in, and finish a multi-step flow. Turning a URL into Markdown, crawling a docs site, or searching the web is a different job, covered in the web scraping guide. The tools below split into a hosted browser fleet, an SDK, a local MCP server, a shell CLI, a self-hosted browser API, and cloud browsers with different bills. We checked every price and free tier on the vendor's own pages on November 3, 2026.
Before you compare prices, decide what the agent actually has to do in the browser:
Click, type, and log in on a page you name. You want a real browser and a way to target elements. Playwright MCP and agent-browser do this on your machine. Stagehand does it in code, locally or on Browserbase. Steel, Kernel, and Browser Use also hand you a cloud browser you can drive with Playwright or CDP.
Hand over the whole flow ("book this", "fill this form", "pull the invoices"). You want an agent loop, not one click at a time. Stagehand's act, extract, and observe, and Browser Use's hosted agent, are built for that. Keep a plain click for the steps that must not wander.
Run it from the coding agent you already have. Add an MCP server or a CLI. Playwright MCP, agent-browser, Browserbase's hosted Stagehand server, Browser Use, and Kernel all have one. Steel's docs show a cookbook for a server you build, rather than a hosted product.
Read, crawl, or search. A browser is the slow, pricey way to turn a URL into text. Use the web scraping guide for URL-to-Markdown, site crawls, and search APIs. Browserbase also sells Fetch and Search beside its browsers; those calls are for context, and the scraping guide compares the tools built for that job.
Vercel Labs also publishes agent-desktop, an Apache-2.0 computer-use CLI that drives desktop apps through the operating system's accessibility tree. That is desktop automation, so it sits outside this browser comparison. Need a model to decide each step? Estimate that bill with our AI API cost calculator.
Tool by tool
Prices are in US dollars, from each vendor's own page on November 3, 2026. Per-hour figures for Kernel are calculated from the per-second rates on its pricing page.
Type · free tier: Hosted browsers for agents · 1 browser hour, 3 concurrent browsers
What it is: a cloud platform that runs fleets of browsers agents can control, plus Fetch and Search APIs, a runtime, and a model gateway. The pricing-page FAQ describes it as headless browsers agents can click through, with Playwright, Puppeteer, Selenium, and Stagehand as the frameworks. Stagehand, the SDK, is separate and MIT.
When to pick it: the agent must log in, click, and finish a flow on a browser you do not want to run, patch, or scale yourself. Paid plans add captcha solving and proxies.
Open source: Stagehand is the MIT piece. The older GitHub repo browserbase/mcp-server-browserbase (Apache-2.0) is archived; its last push was July 20, 2026. The current MCP path is the hosted server below. We did not find a public license for the hosted browser fleet itself.
Pricing:
Free: $0, 3 concurrent browsers, 1 browser hour, 3 Agent runs, 1,000 Search calls, 1,000 Fetch calls, 15 minutes per session, and $5 of model tokens. Captcha solving and stealth mode are off. Data retention is 7 days.
Developer: $20/month, 25 concurrent browsers, 100 browser hours then $0.12 per browser hour, 15 Agent runs, 1,000 Search calls then $5 per 1,000 by default (other search types are priced on the page), 1,000 Fetch calls then $1 per 1,000, and 1 GB of proxies then $12/GB. Captcha solving and basic stealth are on. Data retention is 30 days.
Startup: $99/month, 100 concurrent browsers, 500 browser hours then $0.10 per browser hour, 50 Agent runs, 1,000 Search calls then $5 per 1,000 by default, 10,000 Fetch calls then $0.50 per 1,000, and 5 GB of proxies then $10/GB. Captcha solving and basic stealth are on.
Scale: custom, 250+ concurrent browsers, usage-based hours, advanced stealth, and verified agents plus captcha solving.
MCP: Stagehand's README documents a hosted server at https://mcp.browserbase.com/mcp with a Browserbase API key. It exposes navigate, act, observe, and extract, with no local browser install.
Downsides: one free browser hour and 15-minute free sessions run out as soon as a real flow loops. The free plan has no captcha solving and no stealth mode. Every AI step also spends model tokens. Browserbase's FAQ says it does not offer on-prem browsers. For pages that only need reading, a scrape API is the better bill, which is why that job lives in the scraping guide.
Our take: The default hosted fleet when the agent has to act, and you want the captcha and proxy switches on a paid plan. Pair it with Stagehand. Use plain page commands for steps you can name, and save act for the steps that change.
Type · free tier: Open-source browser SDK (MIT) · the SDK is free; model tokens and the browser are separate
What it is: Browserbase's SDK for browser agents, with TypeScript, Python, and Go. It mixes AI steps (act, extract, observe) with page commands (goto, click, type, locator, screenshot). Docs say it drives Chromium over the Chrome DevTools Protocol, with no Playwright or Puppeteer dependency, and that extract checks results against a schema you define. The GitHub repo has more than 25,000 stars.
When to pick it: you want one script that can click a known selector and also say "open billing" when the DOM moves. It runs on a local Chromium, or on Browserbase.
Open source: MIT, copyright 2026 Browserbase, Inc. Confirmed on the GitHub license API on November 3, 2026.
Pricing: there is no Stagehand subscription. You pay the model provider for act, extract, and observe, and you pay whoever runs the browser. On Browserbase, the free plan includes $5 of model tokens, and the README says you can skip your own model config and use the Model Gateway, with server-side caching for identical calls. A local browser still needs a model key in the quickstart.
MCP: the same hosted Browserbase server as above. There is no separate Stagehand-only cloud fee for that endpoint; it uses your Browserbase key and browser plan.
Downsides: natural-language steps are harder to debug than a selector, and they spend tokens on every call. Browserbase recommends its hosted browsers so you get caching and the Model Gateway, which means a production Stagehand app often picks up a browser-hour bill. The README's login example uses observe so the password stays in your code and off the model; follow that pattern.
Our take: The SDK to start with when you are writing the agent yourself. Run it locally while you learn the loop. Move the same script to Browserbase when you need concurrency, proxies, or captcha solving.
Type · free tier: Official Microsoft Playwright MCP server (Apache-2.0) · free, on your machine
What it is: the MCP server in microsoft/playwright-mcp, published as @playwright/mcp. It lets an LLM open pages, click, type, and fill forms using structured accessibility snapshots, so the model reads roles and text rather than pixels. The latest release is v0.0.84 (October 31, 2026). The repo has more than 39,000 stars. Playwright itself is Apache-2.0.
When to pick it: a coding agent on your laptop (Cursor, VS Code, Claude Code, Claude Desktop, and the other clients listed in the docs) needs to click through a site while you watch. A hosted fleet, residential proxies, or captcha solving from a vendor means Browserbase, Browser Use, Steel, or Kernel instead.
Open source: Apache-2.0, from the GitHub license API. There is no paid plan.
Pricing: no vendor fee. You pay for the machine, the browser, and the model inside your MCP client.
MCP: this is the server. The standard config runs npx @playwright/mcp@latest. Playwright's docs say it is headed by default, and --headless turns that off. You can point it at an existing browser with --cdp-endpoint or --extension. The GitHub README requires Node.js 18 or newer. Playwright's own docs page says Node.js 20 or newer, so match the page you install from.
Downsides: you own proxies, logins, and blocking. The README says coding agents that also hold a large codebase may prefer Playwright CLI with skills, because MCP loads tool schemas and accessibility trees into the context. Docs say browser_run_code_unsafe runs arbitrary Playwright code and is equivalent to remote code execution, so leave it off unless the client is trusted. A headless browser the server launched closes after an hour without tool calls. Headed browsers, and browsers attached over CDP or the extension, stay up unless you set --idle-timeout.
Our take: The first browser to add to a coding agent. It is official, free, and it clicks from the accessibility tree, which keeps runs steadier than screenshot guessing. Move to Browserbase, Browser Use, Steel, or Kernel when the browser has to live in the cloud.
Type · free tier: Open-source browser agent plus cloud browsers · library free; $1 of cloud credit once, no card
What it is: an MIT-licensed browser agent you can run yourself, plus Browser Use Cloud: stealth Chromium at a published browser-hour rate, and a hosted agent API. The README's local paths are a Python library and a CLI. The same README also describes a TypeScript agent and links a separate project, Browser Use Pi. The main repo has more than 119,000 stars. Cloud browsers add stealth, residential proxies, captcha solving, and a live view, per the README.
When to pick it: you want the agent to finish a task ("submit this form") on a hosted browser, or you want the lowest published browser-hour rate and you will watch the proxy line. Use the local library when the browser and the model should stay on your side.
Open source: the library is MIT. The cloud is a paid service.
Pricing: pay as you go, no subscription. Credits never expire.
Cloud browsers: $0.02 per browser-hour. Billing rounds up to whole minutes, with a 1-minute minimum.
Residential proxies: $5/GB, on by default. Direct traffic, or your own proxy, is $0.20/GB. Your proxy vendor can bill you separately.
Hosted web agents: model token rates plus 20%, plus browser time and traffic. A $1 model bill becomes $1.20 before the browser and the proxy. With your own model key, you pay the provider directly and Browser Use still charges 20% of its published token rates, plus browser time and traffic.
Free credit: $1 once per person, for eligible Google, GitHub, or Microsoft signups. No card. The minimum top-up is $5, and the pricing page says that top-up unlocks every model.
Concurrency starts at 10 sessions with $0 paid, then 50 at $100 total paid, 250 at $1,000, 500 at $5,000, and 1,000 at $25,000. Free credits do not count toward those totals.
MCP: a local server (uvx --from 'browser-use[cli]' browser-use --mcp) exposes click, type, navigate, and a fallback agent tool, and the docs call it the free option that needs your own model keys. The cloud server is https://api.browser-use.com/v3/mcp with a bearer API key. Its tools run a task and poll the session. The cloud MCP page says that server runs API V3 sessions, not V4 runs, while Browser Use's current docs tell new integrations to use Cloud API V4.
Downsides: the $1 credit does not renew. A residential proxy at $5/GB, on by default, can cost more than the $0.02 browser hour. The 20% model fee applies on hosted agent runs, including when you bring a key. The V3 MCP versus V4 API split means the MCP path is not the API the docs recommend for new work.
Our take: The pick when you want a cloud browser at a very low hour rate, or a hosted agent that takes a task and returns a result. Turn the default residential proxy off when you do not need it. For a coding agent clicking around locally, Playwright MCP is the simpler add.
Type · free tier: Open-source browser API (Apache-2.0) and a cloud · Launch is $0/month, with $30 of credits once, no card
What it is: a browser API aimed at agents. You create a session and connect Playwright, Puppeteer, or Selenium over CDP. The open-source server runs in Docker. The GitHub repo has more than 7,900 stars. Steel Cloud adds concurrency, proxies, captcha solving, a credentials API, and files.
When to pick it: you want to own one browser on your machine, or you want a cloud session you drive yourself with Playwright. It is a fleet and a session API, not an agent that decides the clicks.
Open source: Apache-2.0. Steel Local is the self-hosted build.
Pricing: browser hours are billed by the minute, rounded up. Docs say the Launch credits are valid for 90 days.
Launch: $0/month plus usage, $30 of credits once (about 300 browser hours if you spend them only on time), 10 concurrent sessions, sessions up to 15 minutes, up to 3 seats, 60 requests a minute, 7-day retention. Browser hours are $0.10. Proxy bandwidth is $10/GB. Captcha solves are $3 per 1,000. Browser tools (/scrape, screenshot, PDF) are $5 per 1,000. No card for the plan. Captcha solving and Steel proxies on Launch need a $10 paid deposit. Free credits do not count toward that deposit.
Scale: $250/month plus usage, $100 of credits every month, 100 concurrent sessions, sessions up to 1 hour, browser hours $0.08, proxies $6/GB, captcha solves $1 per 1,000, dedicated IPs at $5 each, and a HIPAA-ready BAA.
Enterprise: custom, 1,000+ sessions, sessions up to 24 hours. The pricing page includes the Stealth Browser, a custom Chromium fork, on Enterprise only. Launch and Scale list it as not included.
Steel Local, from the local-versus-cloud doc: 1 concurrent session, limited stealth, no captcha solving, bring-your-own proxy, no credentials API, and no files API. Extensions load from a local folder.
MCP: Steel's docs index does not list a hosted MCP product. A cookbook shows how to put a Steel session behind an MCP server you run.
Downsides: Launch sessions stop at 15 minutes, which is short for a login-plus-workflow loop. The $30 is once, and the $10 deposit gates captcha and managed proxies. The pricing page and the local-versus-cloud page describe stealth differently: the comparison page says Cloud stealth is advanced, and the pricing table withholds the Stealth Browser until Enterprise. Read both before you plan on it. Self-hosting is one session, so a fleet means Steel Cloud or your own orchestration. Scale's $250 is a large step up from Launch.
Our take: The open-source browser API to run yourself when one session is enough, and a straightforward cloud when you want Playwright pointed at someone else's Chrome. For an agent that chooses its own clicks, put Stagehand or Browser Use on top. For reading pages in bulk, use the scraping guide.
Type · free tier: Local browser CLI for AI agents (Apache-2.0) · free, no account
What it is: a Rust CLI from Vercel Labs. snapshot prints the page as an accessibility tree with a ref such as @e1 on each element. The agent clicks or types that ref, snapshots again, and repeats. The browser stays open between commands. Chrome for Testing is the default browser. The repo has more than 45,000 stars. Install is npm install -g agent-browser and then agent-browser install, or Homebrew, or npx agent-browser. Binaries ship for macOS (ARM64 and x64), Linux (ARM64 and x64, glibc and musl), and Windows (x64).
When to pick it: the agent already lives in a shell and you want one short command per click, on a browser on that machine. The README also documents attaching to a cloud browser, including Browserbase, Browserless, Browser Use, and Kernel.
Open source: Apache-2.0. There is no paid agent-browser plan. A cloud browser you attach is billed by that vendor.
Pricing: the CLI is free. Building from source, which most people will not do, asks for Node.js 24 or later, pnpm 11 or later, and Rust. Using the published binary does not.
MCP:agent-browser mcp starts an MCP server over stdio. The README also ships skills (agent-browser skills get core) so a harness can load instructions that match the installed CLI.
Downsides: it is a local browser unless you attach one in the cloud, so proxies, captchas, and concurrency are yours. Refs belong to that snapshot; after navigation the README says to snapshot again. By default the daemon closes a headless browser after one hour without commands. This is not agent-desktop, the Vercel Labs CLI for native desktop apps. agent-desktop drives the OS accessibility tree (Finder, Slack, and other apps). It does not replace a page automation tool.
Our take: The local CLI to use when the agent should drive the browser one shell command at a time. Playwright MCP is the closer match for an MCP client that already thinks in tools. Use agent-desktop only when the target is a desktop app, not a web page.
Type · free tier: Hosted Chromium for agents · Developer is free plus usage, with $5 of credits a month and 5 concurrent browsers
What it is: cloud browsers for web agents, with live view, replays, stealth mode, managed auth, and a computer-controls API. The homepage calls it open-source browser infrastructure. The browser image that powers the hosted service is onkernel/kernel-images, Apache-2.0. The cloud plans are paid. You drive sessions with Playwright, CDP, or the hosted MCP server.
When to pick it: browsers sit open between steps and you do not want to pay for idle time. Kernel's pricing FAQ says a browser enters Standby Mode 5 seconds after the last CDP or Live View activity and stops accruing usage cost. Also pick it when you want managed auth (a logged-in profile without pasting passwords into the agent) on every plan, including the free one, with a cap of 3 connections there.
Open source: the browser image repo is Apache-2.0. The hosted control plane is a paid product.
Pricing: included credits apply to usage only. Rates on the pricing page are US dollars per second.
Developer: free plus usage, $5 of credits a month, 5 concurrent browsers, 5 app invocations. Managed stealth, live view, and the computer-controls API are on. Configurable proxies are not. Replays last 1 day. Managed auth connections: 3. Vaults: 3. One project.
Hobbyist: $30/month plus usage, $10 of credits, 10 concurrent browsers, configurable proxies, unlimited managed auth.
Start-Up: $200/month plus usage, $50 of credits, 150 concurrent browsers, bring-your-own proxies, GPU, and regional browsers.
Enterprise: custom concurrency, HIPAA BAA, and the rest of the compliance row on the pricing page.
Usage rates: headless $0.0000166667 per second (about $0.06 an hour), headful $0.0001333336 per second (about $0.48 an hour), headful plus GPU $0.0008000016 per second (about $2.88 an hour). GPU is Start-Up and Enterprise.
Without an account, the pricing page says an agent can buy one stealth headful browser for 30 minutes for $0.50, paid with a Link payment over the machine payments protocol.
MCP: a hosted remote server with OAuth 2.1. Kernel's MCP page says it can launch Chromium, run Playwright, take computer actions, and record replays. The CLI command on that page is kernel mcp install.
Downsides: headful is about eight times the headless rate, so a visible browser changes the bill. Standby stops the usage charge and still holds the concurrency slot until you delete the browser. The free plan has no configurable proxies. The pool FAQ on the same pricing page says pools are billed "per GB-second", while the rate table is headed "Price ($/sec)". Budget from the rate table. The free plan's 5 concurrent browsers sit under Browser Use's 10, and the $5 monthly credit is the whole usage allowance.
Our take: The hosted browser to use when sessions wait on a person or a slow page and idle time should cost zero. Browserbase is the closer match when you want Stagehand and a simple hour bundle. Browser Use is cheaper per active hour if the proxy stays off.
Coding agent on your laptop, today: Playwright MCP. Official, Apache-2.0, and one config away in Cursor, VS Code, or Claude.
The agent already runs shell commands: agent-browser. One snapshot, one ref, one click.
You are writing the automation in TypeScript, Python, or Go: Stagehand. Mix act with a locator you chose.
A cloud fleet with captcha solving and proxies on a paid plan: Browserbase, with Stagehand on top.
The lowest published price per active browser hour: Browser Use at $0.02, and turn off the default $5/GB residential proxy unless you need it.
A browser you run yourself: Steel Local (one session, Apache-2.0) or Playwright on your own machine.
Cloud browsers that should not bill while they wait: Kernel. Standby drops the usage charge after 5 idle seconds.
Read a URL, crawl a site, or search: the web scraping guide. A full browser is the wrong bill for that.
Before you automate
Treat the page as untrusted input. A site can put instructions in the text your agent reads. Playwright's docs also say WebMCP tools registered by the page are untrusted. Keep approvals on for anything that can pay, send, or delete. See our MCP setup guide.
Leave code-execution tools off unless you trust the client. Playwright docs say browser_run_code_unsafe is equivalent to remote code execution. Browser Use's local MCP docs say that server can reach your browser and your file system.
Automate accounts you are allowed to use. Check the site's terms, keep the rate low, and keep passwords out of the model prompt. Stagehand's README uses observe so the fill happens in your code. Kernel's managed auth is the hosted version of that idea.
Cap the spend before the agent loops. Browserbase's free hour is small. Browser Use's residential proxy is on by default at $5/GB. Kernel has organization and project spending caps. Steel Launch will not solve captchas or use Steel proxies until you deposit $10.
Keep keys out of shared configs. Use OAuth where the server offers it (Kernel's hosted MCP does) or an environment variable. Browser Use's cloud MCP docs say to treat the API key as a secret.
FAQ
What's the best browser automation tool for AI agents?
It depends on the job. For a coding agent on your own machine, start with Playwright MCP. For natural-language steps in a real SDK, use Stagehand, on your own Chromium or on Browserbase. For the lowest published cloud browser rate, Browser Use lists $0.02 per browser-hour before proxy traffic. To run the browser yourself, Steel is Apache-2.0.
How is this different from web scraping?
This guide is for acting in a browser: click, type, log in, and finish a multi-step flow. Turning a URL into Markdown, crawling a site, or searching the web is covered in the web scraping guide at https://hotfixdaily.com/guides/web-scraping-for-ai-agents/.
Browserbase or Stagehand?
Stagehand is the MIT-licensed SDK. Browserbase is the hosted browser fleet, and Stagehand's docs recommend running Stagehand there. Use Stagehand on your own Chromium when you want no browser-hour fee. Use Browserbase when you want concurrent cloud browsers, and captcha solving and proxies on a paid plan.
Playwright MCP or agent-browser?
Both are free, local, and Apache-2.0, and both act from an accessibility snapshot. Playwright MCP is the official Playwright server, and the usual add for Cursor, VS Code, and Claude. agent-browser is a Rust CLI with one command per step, which suits a shell-driven agent. Either can attach to a remote browser.
Which of these tools have an MCP server?
Playwright MCP is an MCP server. Browserbase hosts one for Stagehand at mcp.browserbase.com. Browser Use has a local server and a cloud server, and the cloud docs say that server runs API V3 sessions. agent-browser starts one with agent-browser mcp. Kernel hosts a remote server with OAuth. Steel's docs index lists a cookbook for building your own MCP server, not a hosted one.
Do I need a cloud browser?
Use a cloud browser when the agent must run unattended, at higher concurrency, or through a vendor's proxies and captcha solving. A local browser is enough for a coding agent working on your machine. Playwright MCP, agent-browser, Stagehand, the Browser Use library, and Steel Local all run there.
What about desktop apps that are not a browser?
Vercel Labs also publishes agent-desktop, an Apache-2.0 computer-use CLI that drives desktop apps through the operating system accessibility tree. That is desktop automation, so it is outside this browser comparison. For web pages, use agent-browser, Playwright MCP, or a hosted browser.
How we write these: every price, free tier and feature comes from the vendor's own pricing page, docs or GitHub repo (linked as Source) and was checked on the date above. Licenses come from each project's GitHub repo. "Our take" lines are our opinion. Prices change often, so check the link before you pay. Something out of date? Email hello@hotfixdaily.com. No tool paid to be listed. This guide has no affiliate links.
Get the next issue in your inbox
Five things worth your attention in AI, agents, vibe coding and game dev, every weekday. Plus 13 free Starter Prompts and the AI Builder Starter Stack PDF when you join.
