Browser Use
Browser Use is an MIT-licensed open-source AI browser agent (Python and TypeScript) that lets LLMs control a browser, paired with a usage-based cloud offering at $0.02 per browser-hour for managed stealth browsers, CAPTCHA solving, and a hosted agent API.
What is Browser Use?
Browser Use is an open-source browser agent framework, MIT-licensed and available for Python and TypeScript. Its core proposition is making websites accessible to AI agents: instead of writing brittle CSS selectors or XPath expressions, a developer gives the agent a natural-language task, and the LLM decides what to click, type, scroll, or extract on each step by reading a structured DOM representation and, by default, a screenshot of the page.
The product line has three layers. The open-source library is the foundation and can be self-hosted for free. The CLI sits between the library and the cloud for developers who already have a coding agent (Claude Code, Codex, Cursor, Gemini CLI) and want that agent to drive a browser. The cloud offering, Browser Use Cloud, provides managed Chromium browsers with stealth, CAPTCHA solving, residential proxies, and a hosted Agent API (V4) for teams that do not want to run browser infrastructure themselves.
A critical distinction separates Browser Use from browser-automation SaaS platforms: it is a framework for building agents, not a pre-packaged automation product. The official documentation states that for a form that never changes, a deterministic Playwright script is cheaper because it consumes no model tokens. Browser Use is positioned for tasks that vary, where the target site is unfamiliar, or where the workflow cannot be reduced to stable selectors.
How does Browser Use work?
A developer defines a task in natural language and provides an LLM configuration. The agent loop proceeds step by step: it observes the page (structured DOM plus, when use_vision is enabled, a screenshot), sends that observation and the task to the LLM, receives a proposed action (click, type, scroll, extract, navigate), executes the action through Playwright's action API, and repeats until the task is complete or a stop condition is reached. Because the loop calls the LLM on every step, both latency and token cost scale with the number of steps the agent takes.
For self-hosted use, the developer supplies the LLM API key (BYOK) and runs the browser locally or on their own infrastructure. The library supports OpenAI, Anthropic, Gemini, Groq, OpenRouter, Azure, Bedrock, DeepSeek, and local models via Ollama. For cloud use, the developer connects to Browser Use Cloud through the SDK, REST API, or CDP. Cloud browsers add stealth fingerprints, residential proxy rotation, CAPTCHA solving, and a live view URL that allows a human to take over a running session. Session runtime is capped at 240 minutes. The output is whatever the task specifies: extracted data, a completed form confirmation, a file, or a structured JSON object when an output schema is provided.
Use Cases
- Completing forms and portals that have no public API, where the field layout varies across sites or over time
- Extracting data from JavaScript-heavy pages or legacy web applications where selector-based scrapers break
- Running the same multi-step workflow across many different sites that each require their own navigation logic
- Driving a logged-in browser session for tasks that require the user's own cookies, extensions, or authentication state
- Prototyping agentic browser workflows before deciding whether to harden them into deterministic Playwright scripts
- Production automation on bot-protected sites via the cloud stealth layer, with the caveat that CAPTCHA solving is not guaranteed
Pricing
Pricing model: Freemium
| Plan | Price | Billing | Included usage | Notes |
|---|---|---|---|---|
| Free | $0 | one-time credit | One-time $1 credit for eligible Google, GitHub, or Microsoft signups. 10 concurrent sessions. Cost-efficient models only. CAPTCHA solving and webhook events included. No auto-recharge. The credit does not renew. | — |
| Pay as you go | Usage-based | no subscription | $0.02 per browser-hour plus network traffic. Purchased credits never expire. Top-ups from $5 to $100,000. Concurrency scales from 10 to 1,000 concurrent sessions based on total paid: $100 unlocks 50, $1,000 unlocks 250, $5,000 unlocks 500, $25,000 unlocks 1,000. A first top-up unlocks all models. No subscription required. | — |
| Enterprise | Custom | annual | Negotiated annual credit pool. Custom concurrency (1,000+). HIPAA mode, zero data retention on request, EU data residency, SLAs, dedicated support, and a custom data policy. SOC 2 Type II compliant. | — |
Common Limitations
- LLM token cost is unbounded and separate from browser time · The $0.02/browser-hour rate covers only cloud browser runtime. Model tokens are charged from the same credit balance or, for self-hosted use, billed directly by the LLM provider. Browser Use calls the LLM on every step, and with use_vision enabled (the default), each screenshot adds 500–1,500 tokens, which can make a 10-step workflow cost 5–10x more in LLM tokens than a less autonomous, selector-based approach such as Stagehand. For complex pages where the agent scrolls repeatedly to locate an element, token consumption can escalate without a hard ceiling. Disabling vision reduces cost but degrades accuracy on visually complex layouts.
- Cloud privacy default: training on inputs with no documented opt-out · The Browser Use Cloud privacy policy states that user inputs are used to train Browser Use AI models and does not describe a training opt-out for standard accounts. Teams with data-sensitivity requirements must either use the self-hosted MIT library, where the user controls the model provider and data flow, or negotiate a custom data policy with zero data retention on Enterprise. This is not a toggle available on the Free or pay-as-you-go tiers.
- CAPTCHA solving is not guaranteed · Browser Use Cloud includes CAPTCHA solving by default, but the official documentation states that solving 'is not guaranteed.' Complex or unfamiliar CAPTCHA challenges may cause the agent to stop or hand control back to the user. The open-source library, which runs standard Playwright without the cloud stealth layer, is more likely to be blocked by anti-bot systems that detect CDP automation markers. Sites using aggressive bot protection (e.g., Akamai, Cloudflare with strict settings) remain unreliable targets even on the cloud tier.
- Irreversible actions have no built-in rollback · Browser agents operate on live webpages, and actions such as submitting an order, sending an email, confirming a transfer, or deleting an account are not reversible by the tool. Browser Use provides human-in-the-loop checkpoints and live browser takeover, but the agent itself does not enforce confirmation gates unless the user's task or wrapper application explicitly instructs it to stop before irreversible steps. Production deployments that touch consequential actions require additional guardrails designed by the implementing team.
- Self-hosting transfers operational burden to the user · The MIT-licensed library is free, but self-hosting requires managing compute, browser infrastructure, LLM API keys, and deployment. There is no built-in orchestration, no managed stealth or proxy rotation, and no live-view takeover unless the user implements it. Teams without DevOps capacity face meaningful operational overhead compared with fully managed alternatives. The 240-minute session cap applies to Cloud browsers; self-hosted runtimes are bounded by the user's own infrastructure.
Key Features
- Open-source MIT library for Python and TypeScript
- LLM-driven action loop with structured DOM and optional screenshot vision
- BYOK support across OpenAI, Anthropic, Gemini, Groq, OpenRouter, Azure, Bedrock, DeepSeek, and Ollama
- Cloud managed Chromium browsers at $0.02/browser-hour with stealth and residential proxies
- CAPTCHA solving included on cloud browsers (not guaranteed)
Browser Use vs. Stagehand and Browserbase
This comparison uses official product and pricing information available as of November 2026. It focuses on how each product structures the boundary between agent logic, browser infrastructure, and pricing, rather than ranking end-to-end task success on a common benchmark. Browser Use is an agent framework that also sells managed browsers; Stagehand is an SDK; Browserbase is infrastructure. They overlap but are not direct substitutes.
| Tool | Positioning | Key differences | Best fit | Pricing note |
|---|---|---|---|---|
| Browser Use Official product page Machine-readable pricing Open-source library (MIT) | An open-source MIT browser agent framework (Python and TypeScript) with a usage-based cloud offering for managed browsers and hosted agent runs. | The agent is fully autonomous: the LLM decides every step. The library is free to self-host with BYOK across many model providers. Cloud browsers are $0.02/browser-hour plus traffic, with no subscription for new accounts; concurrency is unlocked by total paid rather than by a plan tier. The cloud privacy policy trains on inputs with no documented opt-out. Session runtime is capped at 240 minutes. | Developers and teams that want to own the agent logic, can manage LLM configuration and error handling, and need an autonomous agent for unfamiliar or variable sites—not a deterministic script for a fixed workflow. | Open source: free (MIT), BYOK and self-managed infrastructure. Cloud: $0.02/browser-hour plus proxy traffic; no subscription; non-expiring credits; concurrency from 10 to 1,000 by total paid. Enterprise: custom with HIPAA, zero data retention, and EU residency. |
| Stagehand Official Stagehand site Stagehand vs Browser Use (official comparison) | An MIT-licensed, code-first browser automation SDK from Browserbase that adds LLM-powered primitives to a Playwright-style page object. | Stagehand is deliberately less autonomous than Browser Use. It exposes act(), extract(), and observe() methods that can be mixed with deterministic Playwright code, so the LLM is used selectively rather than on every step. The official comparison states that 'Stagehand is a kit for building an agent; browser-use is an agent someone already built.' For a 10-step workflow, Browser Use can consume 5–10x more LLM tokens than Stagehand. Stagehand is TypeScript-first (Python is also available); Browser Use is Python-first. Stagehand is designed to run on Browserbase, its parent company's infrastructure, but can also run on other CDP-compatible browsers. | Teams that want code precision and selective AI use inside maintainable automations, especially in TypeScript ecosystems, and that prefer Browserbase as the managed browser layer. | Open source MIT SDK, free to self-host (BYOK). Browserbase infrastructure is billed separately from $20/month. No agent-specific token markup is imposed by Stagehand itself. |
| Browserbase Official Browserbase site Browserbase pricing | Managed browser infrastructure for AI agents: headless Chromium sessions, stealth, proxies, and session replay, used as a runtime layer beneath frameworks like Stagehand or Playwright. | Browserbase does not include an agent loop. It sells the browser runtime—CDP endpoints, stealth, proxy rotation, and observability—and expects the developer to bring their own agent logic (via Stagehand, Playwright, or another framework). Browser Use Cloud competes with Browserbase at the infrastructure layer but also sells a hosted Agent API (V4) that runs the agent for you. Browserbase is the parent of Stagehand; Browser Use is framework and infrastructure under one roof, with the framework being open source and the infrastructure being usage-based. | Teams that already have agent logic (or are building it with Stagehand/Playwright) and want a managed browser runtime without adopting a specific agent framework's execution model. | Starts at $20/month for the Developer plan. Usage-based pricing on browser hours and proxies at published rates. No open-source agent framework included; Stagehand is a separate MIT SDK. |
Comparison checked against official product information on 2026-11-03. Features and pricing may change.
Frequently Asked Questions
Is Browser Use free?
The open-source Python and TypeScript library is MIT-licensed and free to self-host, but you pay your own LLM API costs and manage your own browser infrastructure. Browser Use Cloud offers a Free tier with a one-time $1 credit, 10 concurrent sessions, and cost-efficient models only. Beyond that, cloud usage is pay-as-you-go at $0.02 per browser-hour plus network traffic; there are no recurring subscriptions for new accounts.
How much does Browser Use actually cost to run?
Cloud browser time is $0.02 per browser-hour, plus proxy traffic at $5/GB for managed residential proxies or $0.20/GB for direct/ BYO-proxy egress. Model tokens are charged separately from the same credit balance (or billed by your LLM provider if you use BYOK). Because the agent calls the LLM on every step and uses screenshots by default, LLM token cost often exceeds browser time. For self-hosted use, the software is free but you bear LLM API costs and infrastructure costs directly.
Does Browser Use work on sites with CAPTCHAs and bot protection?
Browser Use Cloud includes CAPTCHA solving and a stealth layer, and the company publishes benchmark results claiming a 90.5% pass rate on its Stealth Bench V2 set of bot-protected sites. However, the official documentation states that CAPTCHA solving 'is not guaranteed,' and sites with aggressive bot protection such as Akamai or strict Cloudflare configurations can still block the agent. The open-source library, which runs standard Playwright, is more vulnerable to detection than the cloud stealth browsers.
Does Browser Use train on my data?
The Browser Use Cloud privacy policy states that user inputs are used to train Browser Use AI models and does not describe a training opt-out for standard accounts. If you need to avoid this, use the self-hosted MIT library (where you control the model provider and data flow) or negotiate a custom data policy with zero data retention on Enterprise. Self-hosted telemetry can be disabled with ANONYMIZED_TELEMETRY=false.
What is the difference between Browser Use and Playwright?
Playwright is a deterministic browser-automation framework where a developer writes selectors and action sequences. Browser Use runs on top of Playwright but adds an LLM-driven decision loop: instead of executing fixed selectors, the agent observes the page and decides the next action at runtime. This makes Browser Use more resilient to layout changes and usable on unfamiliar sites, but it adds LLM token cost and latency per step. For stable, repetitive workflows, a plain Playwright script is cheaper and more predictable.
Alternative Tools
-
Devin Devin is an autonomous AI software engineer that plans, writes, tests, and ships code in asynchronous cloud sessions. It suits scoped backlog tasks and feature implementation on bounded codebases, but requires human supervision, struggles on large repositories, and carries a cost profile that is difficult to predict beyond included quotas. -
Genspark Genspark is an all-in-one AI workspace that combines autonomous agents, deep research, AI Slides, Sheets, and Docs, and 70+ models to turn prompts into finished work products—not just chat replies. -
Manus Manus is an autonomous general-purpose AI agent that plans and executes multi-step tasks in a cloud sandbox, producing research reports, websites, slides, data analyses, and working prototypes rather than conversational answers. -
Grok Bot Grok Bot is xAI's AI teammate that runs tasks asynchronously in a cloud computer with browser, file system, and terminal access, available exclusively through The X Subscription's Premium tier and above. -
OpenClaw OpenClaw is an open-source, self-hosted AI agent that runs inside messaging apps like WhatsApp, Telegram, and Slack, executing real tasks such as managing files, sending emails, and automating workflows with your own model API keys. -
LangChain LangChain is an MIT-licensed framework for building agents and LLM-powered applications, with create_agent as its current entry point and LangGraph as the underlying orchestration runtime. LangSmith, a separate paid product, adds tracing, evaluation, and deployment.