Browser Use
Browser Use · Browser Use: Open-Source Browser Agent Library, CLI, and Cloud Browser API
Browser Use is an MIT-licensed Python library and hosted cloud service that lets AI agents operate a real web browser: open pages, click, type, fill forms, and extract data from a natural-language task. It suits developers who want to add web automation to their own agents or apps, either self-hosted with any LLM or on Browser Use's pay-as-you-go stealth browsers.
Best for
Python developers and AI teams who need an agent to operate real websites, either by embedding the open-source library in their own stack or by renting hosted stealth browsers and agents on usage-based pricing
Not ideal for
Non-technical users who want a finished consumer assistant, teams that need a fixed monthly budget for heavy browsing, and scraping jobs that a plain HTTP client or site API can already handle
Who it's for
Python developers, AI agent builders, and automation teams that need agents to interact with live websites
Capabilities
- MIT-licensed Python library with an `Agent` class that turns a natural-language task into browser actions on a local or cloud browser
- Works with 15+ LLM providers in the library, including OpenAI, Anthropic, Google Gemini and Vertex AI, Azure OpenAI, AWS Bedrock, Groq, DeepSeek, Mistral, OpenRouter, LiteLLM, and local models through Ollama
- BU 2.0 (`ChatBrowserUse`), Browser Use's own model tuned for browser automation, available through a Browser Use API key
- Custom tools registered with a `Tools` decorator, structured output, and options to extend or override the default system prompt
- Browser Use CLI and installable skill that give Claude Code, Codex, OpenClaw, Hermes Agent, and other coding agents control of your real Chrome or a cloud browser
- Hosted V4 Agent API with sessions, persistent workspaces for files, rerunnable scripts, scheduled agents, human takeover through a live browser, and Python and TypeScript SDKs
- Hosted stealth Chromium browsers you can drive directly over CDP from Playwright, Puppeteer, or Selenium
- Cloud browsers include fingerprint randomization, residential proxies in 195+ countries (on by default), automatic CAPTCHA solving, live preview, and recordings
- Browser profiles and cookie sync from local Chrome, plus 1Password autofill, TOTP 2FA, and run-scoped secrets for logged-in tasks
- Remote MCP server for Claude Code, Claude Desktop, Cursor, Windsurf, and other MCP clients
- EU deployment in Frankfurt with separate workload storage and a shared account and credit balance
Limitations
- Stealth browsers, residential proxies, and CAPTCHA solving are Cloud features. The self-hosted library drives an ordinary local browser unless you connect it to a paid cloud browser
- CAPTCHA solving and anti-bot bypass are not guaranteed. The docs say a site can still block the browser or require verification
- Cloud cost combines model tokens (plus a 20% fee), browser minutes, and proxy traffic, which makes long or repeated runs harder to budget than a flat plan
- Cloud browsers keep billing until they stop or time out. A finished run or closed CDP connection does not stop the browser immediately
- The agent's own success flag is self-reported. The docs advise verifying important external actions independently
- The open-source library, the Cloud SDK, and Cloud API V2, V3, and V4 have different interfaces, so code examples are not interchangeable
- Profile sync moves saved logins but not extensions, and sites can expire or reject a moved session, so some need a fresh sign-in
- The EU deployment supports a smaller feature set and offers GPT-5.6 Luna and GPT-5.6 Sol as managed models
- Jev Ultrafast is an MVP that needs a TypeSafe API key and does not handle shadow roots, frames, canvas, uploads, or pop-up tabs. SUPERFAST mode for Cloud is waitlist-only
- Accuracy and stealth benchmark figures on the Browser Use site are run and published by the vendor
Use cases
- Adding a browser tool to a Python agent so it can search, fill forms, and extract data from sites without an API
- Giving Claude Code or Codex control of your logged-in Chrome to test a local web app or gather information
- Moving existing Playwright or Puppeteer scrapers onto hosted stealth browsers with residential proxies
- Sending scheduled data-extraction or monitoring tasks to the hosted V4 Agent API and collecting structured JSON results
- Running many browser sessions in parallel for research or lead-list building without managing browser servers
Our take
Browser Use covers more ways to give an agent a real browser than most tools listed here: the same project spans a self-hostable MIT library, a CLI for coding agents, and a hosted browser and agent service. That flexibility is also the main source of confusion: the library, Cloud SDK, and API versions differ, and the features that matter most on hostile sites, such as stealth, proxies, and CAPTCHA solving, sit in the paid Cloud. For teams that only need a remote browser, the $0.02 browser-hour is simple to reason about. Hosted agent runs are harder to budget because token, browser, and traffic costs stack, so it is worth measuring a representative task before scaling it. Treat the site's benchmark numbers as vendor-reported.
Who should use it
Developers building agents that must use websites without APIs, teams that want to give Claude Code or Codex a browser, and automation teams that want hosted stealth browsers for existing Playwright or Puppeteer scripts.
Who should skip it
People looking for a ready-made consumer assistant with no code, teams that cannot accept variable usage costs, and workloads where a site API or simple HTTP scraping is enough.
Strengths
- MIT-licensed library you can self-host with any of 15+ model providers, including local models
- One project covers a Python library, a CLI for coding agents, hosted browsers, and a hosted agent API
- Hosted browsers work with existing Playwright, Puppeteer, and Selenium code over CDP
- Pay-as-you-go credits with no subscription and no expiry, plus $15 in free credits for eligible signups
- Stealth, proxies, and CAPTCHA solving are included in the browser price rather than sold as add-ons
Weaknesses
- Anti-bot features require the paid Cloud
- Usage costs across tokens, browser minutes, and traffic can be hard to predict
- Several API generations and separate SDKs make it easy to mix up examples
- The main agent library is Python. JavaScript and TypeScript users rely on the Cloud SDK or separate projects such as Browser Harness JS
Browser Use pricing
Open-source library
Free
- MIT license
- Self-host with your own model and browser
- Model provider usage billed by that provider
Browsers (Browser Infrastructure)
$0.02 per browser-hour
- Residential proxies $5/GB, on by default
- Direct or your own proxy $0.20/GB
- Billed by the minute, 1-minute minimum
Browser Use Agents
Model cost + 20%
- Plus browser time and traffic at browser rates
- Example: GPT-5.6 Luna $0.24 input / $1.44 output per 1M tokens, fee included
- Bring your own Anthropic, OpenAI, or Google key and pay 20% of published token rates
Custom model API (BU 2.0)
$0.60 input / $3.50 output per 1M tokens
- Cached input $0.06 per 1M tokens
- Use Browser Use's model with your own agent
Enterprise
Custom
- Custom concurrency and SLAs
- Custom data policy, HIPAA support, and zero data retention
Free tier limits: $15 in one-time free credits for eligible Google, GitHub, or Microsoft signups, with no card required. Projects start at 10 concurrent sessions.
Note: No subscription plans are offered. Credits start at a $5 minimum top-up and never expire. Concurrency is shared by agents and browsers and rises with lifetime paid spend (free credits do not count), from 10 sessions up to 1,000 at $25,000. The pricing page and the concurrency docs currently disagree on the 50-session threshold ($200 vs $100); the docs say to treat the account's own limit as the source of truth. Legacy V2 and V3 API pricing is separate.
Technical specs
Modalities
Web browser (Chromium/Chrome), Text
API pricing
Pay-as-you-go credits: $0.02 per browser-hour plus proxy traffic; hosted agents add model tokens plus a 20% fee
Available models
Where Browser Use excels
Browser tool for a custom agent
The Python library adds web navigation, form filling, and extraction to an existing agent, with the model and browser under your control.
Coding agent that checks a web app
The CLI skill lets Claude Code or Codex open your local Chrome with its logins and inspect or exercise a running app.
Scraping protected sites at scale
Existing Playwright or Puppeteer code connects over CDP to hosted browsers with residential proxies and CAPTCHA solving, billed per minute.
Browser Use vs. competitors
Browser Use vs. Kitesurf
Kitesurf is Cloudflare's Chromium-free browser runtime used through Browser Run's CDP endpoints, while Browser Use Cloud hosts stealth Chromium browsers with proxies and CAPTCHA solving and also offers a hosted agent on top.
Browser Use vs. Manus
Manus is a general task-execution agent you use through its own app, while Browser Use is a browser-specific agent you call from code, a CLI, or an API, and can self-host as an MIT library.
Browser Use vs. Smolagents
smolagents is a general-purpose Python library for code-writing agents, while Browser Use is a Python library built around controlling a browser, with an optional hosted browser and agent service.
Browser Use vs. Momentic Mo
Momentic Mo is a hosted QA product that explores an app and files the bugs it reproduces, while Browser Use is general browser automation you script for any task, including testing, without built-in bug reporting.
Frequently asked questions
What is Browser Use?
Browser Use is an open-source Python library and a hosted cloud service for AI agents that operate a web browser. You give an agent a task in plain language and it navigates, clicks, types, and extracts data on real websites.
Is Browser Use open source?
Yes. The browser-use library on GitHub is MIT-licensed and can be self-hosted with your own model and browser. Browser Use Cloud, which adds hosted stealth browsers and a hosted agent API, is a paid service.
How much does Browser Use cost?
The library is free. Browser Use Cloud is pay-as-you-go: browsers cost $0.02 per hour plus $5/GB for residential proxy traffic ($0.20/GB direct or with your own proxy), and hosted agents cost model tokens plus a 20% fee on top of browser time. Credits start at $5, never expire, and eligible new signups get $15 free. Enterprise pricing is custom.
Which LLMs does Browser Use support?
The open-source library supports 15+ providers, including OpenAI, Anthropic, Google, Azure OpenAI, AWS Bedrock, Groq, DeepSeek, Mistral, OpenRouter, LiteLLM, Ollama, and any OpenAI-compatible endpoint, plus Browser Use's own BU 2.0 model. The hosted Cloud V4 API defaults to GPT-5.6 Luna and also accepts models from Anthropic, OpenAI, Google, xAI, Z.ai, Moonshot AI, and MiniMax, though the docs note that the dashboard shows a smaller set and access can depend on your account.
Can I use Browser Use with Claude Code or Cursor?
Yes. The Browser Use CLI installs a skill that lets Claude Code, Codex, and other coding agents control your Chrome or a cloud browser, and the hosted MCP server connects to Claude Code, Claude Desktop, Cursor, and Windsurf with a Browser Use API key.
What is Jev Ultrafast?
Jev Ultrafast is an experimental, MIT-licensed browser agent that Browser Use open-sourced in September 2026. It uses TypeSafe's Jev model to choose an action and a page element in one request and calls a small LLM only when it needs to type text. It requires a TypeSafe API key and is a separate repository from the main browser-use library.
Integrations & fit
Alternatives to consider
Cloudflare
Developers and platform teams running web-browsing AI agents at scale on Cloudflare who want cheaper, higher-concurrency sessions
Free
Manus
Professionals and teams who need AI to complete defined tasks end-to-end — research reports, data extraction, scheduled automation, presentations — rather than assist with each step manually
FreemiumHugging Face
Minimal code-first agents with Hugging Face ecosystem
FreeAbout Browser Use
Browser Use offers three ways in, and the right one depends on who runs the agent. The open-source Python library (Python 3.11+, `uv add browser-use`) gives you an `Agent` class that takes a task and an LLM and drives a local Chromium or Chrome, with custom tools, structured output, and your choice of 15+ model providers, including local models through Ollama. The Browser Use CLI, backed by the separate Browser Harness project, is aimed at coding agents such as Claude Code, Codex, Hermes Agent, and OpenClaw: you install a skill and the agent controls your real Chrome, with its tabs and logins, or a cloud browser. Browser Use Cloud is the commercial side. It has two products: Browser Use Agents, where you send a goal to the V4 API and Browser Use runs both the agent and the browser, and Browser Infrastructure, which hosts stealth Chromium browsers that your own Playwright, Puppeteer, Selenium, or CDP code connects to. Cloud browsers include fingerprint randomization, residential proxies, automatic CAPTCHA solving, persistent profiles, live preview, and recordings, and there is a remote MCP server for Claude, Cursor, and Windsurf. Pricing is pure pay-as-you-go: $0.02 per browser-hour plus proxy traffic, and model tokens plus a 20% fee for hosted agents, with credits that never expire and no subscription. In September 2026 the company also open-sourced Jev Ultrafast, an experimental MIT-licensed agent that uses a TypeSafe model to pick an operation and a target element in one request, and its site now takes waitlist sign-ups for a SUPERFAST mode in Cloud. The tradeoffs: the library itself is free, but stealth, proxies, and CAPTCHA handling live in the paid Cloud; costs combine tokens, browser minutes, and traffic, which makes long runs harder to budget; and the open-source library, Cloud SDK, and three generations of Cloud API (V2, V3, V4) have different interfaces, so examples are not interchangeable.
Updates from Browser Use
Browser Use published Jev Ultrafast, an MIT-licensed experimental browser agent built with TypeSafe that chooses an operation and a target element in a single model request and calls a small LLM only to type text. The repository points to a waitlist for early access to ultrafast agents in Browser Use Cloud, which the site calls SUPERFAST mode.
The library release upgraded Browser Harness, pinned dependencies, migrated to MCP Python SDK 2.1.1, and resolved three pypdf security advisories.
V4 agents can now be scheduled with pause and resume controls, and a new Agency skill lets agents prepare work and ask for approval before external actions such as sending or publishing.
Are you the founder? Claim this listing →