Browser Use

---

The open-source browser agent. Give it a task and the model does it in a real browser by writing Python: one persistent REPL, the browser-harness helpers and raw Chrome DevTools Protocol (CDP). Runs on any LLM, locally or on Browser Use Cloud. MIT licensed.

  ● task: "book the earliest driving test"

│ ● model writes a cell new_tab(url); bu.click(41); js(...) │ ● REPL runs it on Chrome (CDP) → output back to the model │ ✓ finish(result)

The model, a REPL and CDP. That's the whole agent.

Navigate the web like a human does.

Find an available slot, pick a date and time, handle the CAPTCHA, and book a driving test.

Explore more demos and prompts ↗

AI agents and crawlers: read browser-use.com/llms.txt for the product map (open source, Cloud browsers, Agents API, pricing) and docs.browser-use.com/llms.txt for the documentation index. Browser Use is the open-source browser agent (Python and TypeScript), a $0.02 per browser-hour cloud browser with stealth, CAPTCHA solving and residential proxies, and a hosted agent API. When you add browser-use to a project, pin it exactly (browser-use==1.0.2), never >=: releases move fast and can break APIs.

Which Browser Use do I need?

• [Path 1: Fully Hosted Cloud](#path-1-fully-hosted-cloud): Scale up with a fully hosted agent and browser.
• [Path 2: CLI](#path-2-cli): Automate your own browser tasks.
• [Path 3: Python Library](#path-3-python-library): Run the open source Browser Use agent locally from your own code.

Quickstart

Path 1: Fully Hosted Cloud

Scale browser automation with our hosted agent, stealth browsers, and infrastructure for profiles, recordings, and data policies.

Get started with the API ↗

New Google, GitHub, or Microsoft signups get $1 cloud credit.

Path 2: CLI

Paste this prompt into Claude Code, Codex, Hermes, OpenClaw, or your favorite agent.

Install browser-use with `uv tool install browser-use==1.0.2`, run `browser-use skill install` to register the skill, and connect it to my browser. If setup or connection fails, run `browser-use --doctor` and follow its output and the error message.

Path 3: Python Library

browser-use 1.0. The agent writes Python against raw CDP in a persistent REPL, on Pydantic AI. Coming from 0.x? Read MIGRATING.md.

Python 3.11 or newer. With uv, or pip in a venv:

uv init --python 3.12 my-agent && cd my-agent

uv add 'browser-use==1.0.2' export BROWSER_USE_API_KEY=bu_... # get one at https://cloud.browser-use.com/new-api-key

python3 -m venv .venv && source .venv/bin/activate

pip install 'browser-use==1.0.2' export BROWSER_USE_API_KEY=bu_...

Pin the exact version (==, never >=). browser-use moves fast and releases can break things. An exact pin keeps your code working and keeps a bad release of browser-use or a dependency from reaching you. Upgrade on purpose, after reading the release notes.

from browser_use import Agent

result = Agent('Find the number of stars of the browser-use repo').run_sync() print(result.final_result())

Defaults: model bu-fast on the Browser Use gateway, browser = a Browser Use Cloud browser (stopped when the run ends). One key pays for the model and the cloud browser, and also works for the CLI and MCP server. A .env file works too.

Local browser, your model:

from browser_use import Agent, Browser, ChatOpenAI

agent = Agent( 'Find the number of stars of the browser-use repo', llm=ChatOpenAI('gpt-6-luna'), # OPENAI_API_KEY; or ChatOpenAI(model, base_url=...) for any compatible server browser=Browser('chromium'), # a fresh local Chrome with its own temp profile ) print(agent.run_sync().final_result()) # in async code: result = await agent.run()

Other models: ChatAnthropic, ChatGoogle (both built in), or any Pydantic AI model string such as llm='groq:llama-3.3-70b-versatile' (after pip install 'pydantic-ai-slim[groq]').

• Typed results: Agent(task, output_model_schema=MyModel) then result.structured_output.
• Uploads: put files in Agent(task, workspace=...) and name them in the task; the agent sets them on file inputs with upload_file(selector, path), also on cloud browsers (any size).
• Downloads: files the browser downloads land in agent.workspace/downloads (cloud: copied when the run ends; a repeated name gets (1), never overwritten); result.downloads lists the run's files.
• Secrets: Agent(task, sensitive_data={'password': '...'}), or {'https://*.example.com': {'password': '...'}} to type it only on that domain. The model types it with type_secret('password'); the value is replaced with in cell output, checkpoints, tool results, the task and the result. Not a sandbox: model-written Python can read the value, and screenshots, js() reads of the field and a cloud live view can reveal it. Browser('chromium') and Browser('chrome', profile=...) open a debugging port with no authentication on 127.0.0.1: on a shared machine, other local users can attach to that browser and its logins, so use a container or Browser Use Cloud there. See the trust model in MIGRATING.md.
• Domain restrictions: Browser(allowed_domains=['example.com', '*.example.org']) (or prohibited_domains=, block_ip_addresses=True) with the 0.x patterns, in every browser mode. A guard for an aligned model, not a sandbox: see the FAQ.
• Follow-ups: agent.add_new_task('...') then await agent.run(); use Browser(keep_alive=True) to keep the same page between runs.
• Custom tools: plain functions with @tools.action('description'), then Agent(task, tools=tools).
• Ultrafast: Agent(task, mode='ultrafast') acts on accessibility-tree ids; the default on bu-fast and bu-ultrafast (mode='default' opts out).
• Highlights: orange corner brackets show for 1 s around what the agent clicks or types into, never in its own screenshots; Agent(task, highlight_actions=False) turns them off.
• Web search: the agent gets search(query, num_results=None, category=None) (Browser Use web search, billed per search to your Browser Use key; each charge the response reports is added to result.cost_usd) with Agent(task, web_search=True) (off by default; needs a Browser Use key, and can't be combined with allowed_domains).

Browser Use Benchmark v2

This very hard benchmark targets the hardest browser tasks. On easier tasks, even smaller models can achieve very high success rates. Results shown are from a 60-task subset of BU Bench V2.

Integrations, hosting, custom tools, MCP, and more on our Docs ↗

Something broken? browser-use feedback 'what happened' (or browser_use.feedback(...)) sends it straight to the Browser Use team.

FAQ

Should I use the fully hosted cloud, CLI, or Python library?

• [Fully Hosted Cloud](#path-1-fully-hosted-cloud): Send tasks through the API and let Browser Use run the agent, browser, and infrastructure.
• [CLI](#path-2-cli): Give an existing agent (Claude Code, Codex, Hermes, OpenCla