Browser Use Review

Browser Use is a developer-facing library for agents that use the browser. The OpenAlternative page lists multi-tab sessions, DOM extraction that gives the agent a structured view of the page, vision for visually dependent pages, parallel execution across browser instances, and session persistence including cookies and logins. The README adds custom tools registered with a decorator, custom system prompts, support for many models through a gateway or provider wrappers, a CLI with a skill install for coding agents such as Claude Code or Codex, and a Claude integration that exposes 31 browser actions plus Bash. You can run a local browser with a local model through Ollama, or use the hosted cloud, which offers a hosted agent API and cloud browsers with stealth features, with a BU2 model available through an API key.

Python 3.11 or later is required, with 3.12 in the examples. In a new project you run uv init --python 3.12 and uv add browser-use, put a model API key such as OPENAI_API_KEY in .env, and write a short agent.py that creates an Agent with a task and an LLM and runs it with uv run agent.py. A BROWSER_USE_API_KEY is optional and enables the BU2 model or cloud browsers. For the hosted Cloud API V4 the Python SDK needs Python 3.10 or newer (pip install browser-use-sdk playwright), you connect over a CDP URL, and you must stop browsers explicitly because disconnecting does not stop them. The Bash tool needs Linux or macOS, and Windows users need WSL.

The experience is code: a task string and a model produce a run, and the agent opens a browser, searches, and returns an answer. The hosted API charges pay as you go, enforces project rate limits in five-second windows, and returns 409 if a session is busy, and a client-side wait timeout does not cancel the server-side run. This section reflects the README and docs; we did not run an agent ourselves.

• Multi-tab sessions and structured DOM extraction
• Vision support for visually dependent pages
• Parallel agents across many browser instances
• Session persistence with cookies and logins
• Custom tools, custom system prompts, and many model providers
• Local mode with a local browser and Ollama, or a hosted agent API and stealth cloud browsers
• CLI and skill installation for coding agents

• The most popular open-source browser-agent library, with about 118,000 stars
• Fully usable offline-style with local models, or scaled up with the hosted cloud
• Model-agnostic, with a short API that is easy to prototype with
• Clear separation between the free library and the paid hosted services

• A code library with no low-code interface, so it needs a developer
• No configuration guarantees that every CAPTCHA can be avoided or solved
• Profile sync transfers cookies only, not local storage, IndexedDB, or extensions, so some sites require signing in again
• Local models depend on your hardware, and Bash needs Linux, macOS, or WSL
• The README cites benchmark results on a 60-task subset, so treat accuracy claims as indicative, not guaranteed

The library is MIT-licensed and free. Model inference and hosted browsers are billed separately: Browser Use Cloud lists cloud browsers at $0.02 per browser-hour with a $15 credit for new Google, GitHub, or Microsoft signups, and its API billing is pay as you go. You can avoid the cloud entirely with a local browser and a local model through Ollama.

Browserbase is cloud headless-browser infrastructure for AI agents and automation, covering sessions, scaling, and anti-detection. Browser Use is the agent layer, and it can run on your own browser or on remote ones, including its own hosted stealth browsers. The two can be combined: Browser Use for the agent logic and a hosted browser for the sessions. Choose Browserbase when you only need managed browser sessions for your own Playwright or Puppeteer code, and Browser Use when you want the agent loop included.

• Developers building agents that must operate real websites, including login flows and forms
• Teams that want to start free with local models and scale to hosted browsers later
• Prototypes for research, data entry, and QA agents
• Anyone who wants to give a coding agent browser control

• Teams that want a no-code tool
• Large-scale scraping where a dedicated crawler is cheaper and more reliable
• Targets with strong bot protection, unless you accept the CAPTCHA limitations