Browser Use

Drive a real browser from a Python agent and execute multi-step web tasks

The official README presents Browser Use as an AI browser agent. With the Python library, an Agent receives a natural-language task and an LLM, then operates a browser to complete the task. Developers can extend the workflow with custom tools and structured output for application-specific automation.

Sources: [1]

Choose CLI, Python library, or hosted cloud based on how much infrastructure you want to manage

The README separates three usage paths. The CLI gives an existing agent such as Claude Code or Codex browser access, the Python library runs your own agent code, and Fully Hosted Cloud runs both the agent and browser through an API. Both the CLI and Python library can connect to either a local or cloud browser.

Sources: [1]

Choose the LLM provider and browser target independently, from local to managed setups

Browser Use supports direct provider wrappers such as OpenAI, Anthropic, and Google, the Browser Use model gateway, and local model options. Browser execution can also be local or cloud-hosted, so model inference, credentials, and browser infrastructure do not have to be tied to one vendor.

Sources: [1][3]

Package browser automation and agent tooling as a Python 3.11+ library with CLI entry points

The repository topics include playwright and browser-automation, while the package requires Python 3.11 or newer. Its pyproject declares dependencies for OpenAI, Anthropic, Groq, Ollama, MCP, and browser-harness, and exposes browser-use as a CLI entry point.

Sources: [2][3]

A strong fit when an agent must operate human-oriented web interfaces that do not expose the API you need

Browser Use is useful for research, booking, form entry, and other browser tasks that need to become part of an agent workflow. Teams can prototype with the Python library and a local browser, then move to managed browsers or a hosted agent only when scale, stealth, profiles, or operations justify it.

Browser automation carries account, cookie, and real-world side effects, so permissions need explicit boundaries

When a local Chrome profile is reused, the agent can access sites where that profile is already authenticated. Cloud profile sync transfers cookies but not local storage, IndexedDB, or extensions. Tasks that can write, purchase, book, or publish should use dedicated profiles, least-privilege accounts, confirmation boundaries, and auditable logging.

Cloud browsers provide managed stealth, proxies, profiles, and recordings, while Fully Hosted Cloud also runs the agent on Browser Use infrastructure. The open-source Python library is MIT-licensed, but model providers and Browser Use Cloud are separate paid services with their own terms. Review what data is sent to the LLM provider, browser infrastructure, and hosted agent for each deployment.

Sources: [1][4]