An AI agent that has to use a website needs three things: a browser, a way to read and act on the page, and a loop that decides what to do next. Browser Use, Browserbase and Stagehand each supply a different part of that stack, and Steel, Hyperbrowser and Playwright MCP fill gaps around them. This guide scores six tools for developers building their own agents, with pricing as published in October and checked on November 3, 2026.
Which layer of the stack do you need?
Browser infrastructure for agents splits into three layers, and most confusion in this market comes from comparing tools across them. An agent library (Browser Use) takes a goal and plans every click. An automation SDK (Stagehand) lets you script the workflow and call a model only for steps a selector cannot handle. A hosted browser (Browserbase, Steel, Hyperbrowser) runs Chromium in the cloud with sessions, proxies and CAPTCHA handling, and does not care which of the first two drives it. Playwright MCP is a fourth shape: it turns a local browser into tools any MCP client, such as a coding agent, can call.
The layer decides the failure mode. A goal-driven agent handles pages it has never seen and fails in ways that are hard to predict. A scripted workflow with AI steps fails less and costs fewer tokens, but someone has to write it. A hosted browser fixes neither; it removes the operational work of running many isolated, logged-in browsers. Pick the layer first. The vendor second.
End-user agentic browsers such as Comet or Claude in Chrome are a different purchase, covered in Best AI Browsers.
How do the tools score?
Four criteria, 1 to 5 each, from vendor documentation, pricing pages and release notes through November 3, 2026. These are editorial assessments of published capabilities and prices, not benchmark runs.
- Agent: how much of the agent loop the product supplies, from single AI steps to full goal-driven runs.
- Infrastructure: hosted browsers, persistent sessions and logins, stealth, CAPTCHA and proxies, replay, concurrency.
- Openness: license, self-hosting, choice of model and driver, ease of leaving.
- Cost: the published price of getting started and the unit rate at volume.
| Tool | Layer | Agent | Infra | Openness | Cost | Total /20 | Published price |
|---|---|---|---|---|---|---|---|
| Browser Use | Agent library + cloud | 5 | 4 | 5 | 5 | 19 | Library free (MIT); cloud $0.02/browser-hr, steps from $0.006 |
| Stagehand | Automation SDK | 4 | 2 | 5 | 5 | 16 | Free (MIT); pay for model and browser |
| Browserbase | Hosted browsers + Agents API | 4 | 5 | 3 | 3 | 15 | Free 1 hr; $20/mo (100 hrs, then $0.12/hr); $99/mo (500 hrs, then $0.10/hr) |
| Steel | Browser API (self-host or cloud) | 2 | 4 | 5 | 4 | 15 | Hobby $0 ($10 credit); $29 to $499/mo; $0.10 to $0.05/hr |
| Playwright MCP | MCP server (local) | 3 | 1 | 5 | 5 | 14 | Free (Apache-2.0) |
| Hyperbrowser | Hosted browsers + hosted agents | 3 | 4 | 2 | 4 | 13 | $0.10/browser-hr; agent steps $0.02 |
What does each tool offer?
Browser Use
Browser Use is an MIT-licensed Python library that turns a plain-language task into browser actions with any model, including local ones through Ollama (version 0.13.14, October 26, 2026). Its own model, BU 2.0, released January 27, 2026, is trained to choose the next browser action and scored 63.3% on the company's 100-task benchmark. The cloud publishes the lowest browser rate we found: $0.02 per browser-hour, agent steps from $0.006, $0.01 per task and $5 per GB of residential proxy, with SOC 2 Type II and a HIPAA mode on Enterprise (pricing). Best fit: teams that want a goal-driven agent they can run locally first. Full Browser Use review.
Stagehand
Stagehand is Browserbase's MIT-licensed SDK for automation that mixes code with AI calls to act on, extract from and observe a page, plus an agent mode for open-ended tasks. Version 3 (October 29, 2025) dropped Playwright for direct Chrome DevTools Protocol control, added caching of discovered actions and a self-healing layer, and Browserbase reported it 44% faster than v2 on its own suite. Since January 13, 2026 there are SDKs for TypeScript, Python, Go, Ruby, Java and Rust. It runs against a local browser or a Browserbase session. Best fit: engineers who want automation in code review and CI, with a model only where pages are unpredictable. Full Stagehand review.
Browserbase
Browserbase hosts Chromium for agents, each browser in its own VM that is destroyed after the session. Contexts keep logins between runs, CAPTCHA solving starts on the Developer plan, advanced stealth is on Scale, and replay and live view show what an agent did. On June 30, 2026 it launched Browserbase Agents, which runs an agent from a single prompt. Plans as published in October 2026: Free with 1 browser hour, Developer $20 a month with 100 hours then $0.12 an hour, Startup $99 with 500 hours then $0.10 an hour, Scale custom (Toolradar). Best fit: production fleets that need many isolated, logged-in sessions. Full Browserbase review.
Steel
Steel's core, steel-browser, is an Apache-2.0 browser API that runs from a single Docker command, with a separate MIT MCP server (eCorpIT). Steel Cloud adds CAPTCHA solving, proxies and profiles. Published plans: Hobby free with $10 of credit at $0.10 an hour; Starter $29, Developer $99 and Pro $499 a month, with browser time falling to $0.05 an hour on Pro, CAPTCHA solves from $4 to $3 per 1,000 and proxy data from $10 to $5 per GB. Steel supplies no agent of its own. Best fit: teams that want a hosted browser today and the option to self-host later.
Playwright MCP
Microsoft's Playwright MCP server (Apache-2.0) lets any MCP client drive a browser through structured accessibility snapshots rather than screenshots, which keeps it fast and cheap in tokens. It can start each session in an isolated profile, keep a persistent profile with saved logins, or connect through a browser extension to tabs you already have open. It hosts nothing and scales no further than the machine it runs on. At 14,800 US searches a month it is also the most searched tool in this guide. Best fit: coding agents and single-user setups.
Hyperbrowser
Hyperbrowser sells hosted browsers and hosted agents on one credit balance, with 1,000 credits to the dollar. Published rates: browser time $0.10 an hour, proxy data $10 per GB, and $0.02 per step to run its HyperAgent or Browser Use, plus model tokens. Paid plans add stealth, CAPTCHA solving and residential proxies; its HyperAgent framework is AGPL-3.0, and the service itself is not self-hostable. Best fit: teams that want browsers, scraping and hosted agents from one bill.
Also worth knowing
- Kernel publishes about $0.06 an hour for headless browsers and about $0.48 for headful ones, with a standby mode that keeps cookies and page state; its images are Apache-2.0.
- Browserless is self-hostable under SSPL or a commercial license, billed in 30-second units from $25 a month on annual plans.
- Opera Neon is an end-user browser that outside agents can drive for free over MCP or a CLI since August 2026. Opera Neon review.
Kernel and Browserless details from eCorpIT's October 22, 2026 comparison.
How should you choose?
- Write down whether the workflow is known. Known steps belong in Stagehand or plain Playwright with AI fallbacks. Open-ended goals need an agent such as Browser Use.
- Start local. Browser Use, Stagehand and Playwright MCP all run against a browser on your machine. Pay for hosting when concurrency, logins or proxies force it.
- Price the model alongside the browser. Browser time runs from $0.02 to $0.12 an hour, and an agent that re-plans every step also pays for model tokens on each one. Cached actions in Stagehand and step pricing in Browser Use Cloud change that math.
- Check the exit. Steel, Browserless and Kernel publish self-hostable browsers. Code written against standard CDP or Playwright moves between hosts; vendor agent APIs do not.
- Treat page text as untrusted input. An agent that reads a page can be steered by it. Keep agent sessions away from production credentials and log every action.
Related questions
- Browser Use, Browserbase or Stagehand: which fits your agent?
- Which agentic browsers are worth shortlisting for end users?
- When should you use an AI agent instead of scripted automation? (partner site CTAIO)
- What are the security risks of MCP servers? (partner site CTAIO)
- Which orchestration framework should run the agent?
Sources
- Browser Use, Pricing (checked Nov 3, 2026); PyPI, browser-use (Oct 26, 2026); BU 2.0 model card (Jan 27, 2026).
- Browserbase, Launching Stagehand v3 (Oct 29, 2025); Stagehand in every language (Jan 13, 2026); Introducing Browserbase Agents (Jun 30, 2026).
- Toolradar, Browserbase pricing (confirmed Oct 30, 2026); PricingSaaS, Browserbase pricing history (Oct 19, 2026).
- Steel, Pricing (checked Nov 3, 2026).
- Hyperbrowser, Pricing (checked Nov 3, 2026).
- Microsoft, Playwright MCP (GitHub).
- eCorpIT, Cloud browsers for AI agents 2026 (Oct 22, 2026).
Related reading on our partner site CTAIO: agentic browsers, a CTO's policy guide and prompt injection explained.
What is the best browser automation tool for AI agents?
For most teams starting out, Browser Use: the open-source library gives you a complete agent with any model, and its cloud has the lowest published browser rate at $0.02 an hour. For workflows you can script, Stagehand gives more predictable runs. For production fleets that need isolated logged-in sessions, CAPTCHA solving and replay, Browserbase has the most complete hosted platform.
What is the difference between Browser Use, Browserbase and Stagehand?
Browser Use is an agent: you give it a goal and it decides each click. Stagehand is an SDK: you write the workflow in code and call a model for the steps that are hard to script. Browserbase is the hosted browser both can run in. Stagehand and Browserbase come from the same company; Browser Use has its own cloud.
Do I need a cloud browser for an AI agent?
Not to start. Playwright MCP, Browser Use and Stagehand all run a local browser. A cloud browser becomes worth paying for when you need many sessions at once, sessions that keep logins between runs, residential proxies or CAPTCHA solving, or a recording of what the agent did.
How much does a headless browser for AI agents cost?
As published in late October and early November 2026: Browser Use Cloud $0.02 per browser-hour; Steel $0.10 to $0.05 an hour depending on plan; Hyperbrowser $0.10 an hour; Browserbase $20 a month for 100 hours, then $0.12 an hour. Model tokens are billed separately.
Is Playwright MCP enough to give an agent a browser?
For a single developer or a coding agent, often yes. Microsoft's Playwright MCP server is Apache-2.0 and lets any MCP client drive a local browser through accessibility snapshots instead of screenshots. It does not host browsers, manage proxies or scale out, which is where the paid platforms come in.
Ready to Find the Right AI Tools?
Browse our data-driven rankings to find the best AI tools for your team.