Browser Use lets an AI agent operate a web browser the way a person does: it opens pages, clicks buttons, types, and fills in forms. You describe a task in plain language and the agent carries it out — filling a job application from a résumé, extracting structured data about followers and exporting it as CSV, or comparing products across sites.

There are two ways in. The command-line skill installs into an agent you already use — Claude Code, Codex, Cursor, Hermes, OpenClaw, and others — and hands it browser control for one-off tasks. The Python library (Python 3.11 or later) is for building software that automates the web: scheduled or parallel runs for scraping, monitoring, and QA, browser agents embedded in your own product, custom tools, custom system prompts, and structured output.

Model choice is open. The ChatBrowserUse client accepts provider-prefixed model ids, so a single Browser Use API key reaches OpenAI, Anthropic, and Google models without separate provider keys, and you can pass your own provider client or run local models with Ollama instead. An open-source preview model is also available, and the default agent system prompt is sent either way. Custom tools are registered with a decorator on a Tools object and passed to the agent.

The project publishes a benchmark of 100 real-world browser tasks as open source, and reports an 87.4% average on the Odysseys leaderboard of 200 long-horizon web tasks. A hosted cloud service runs the browsers instead of your machine, adding proxy rotation, CAPTCHA handling, stealth fingerprinting, a persistent filesystem and memory, rerunnable scripts that fetch live data as sites change, and an HTTP API for submitting runs.

Features

Natural-language tasks: an agent receives a task description and drives the browser to complete it, with no scripted click path
Structured extraction: page data is exported to CSV and other structured formats
Agent skill: installs into Claude Code, Codex, Cursor, Hermes, OpenClaw, and similar agents through a setup prompt
Python library: an async Agent API for scheduled, parallel, or embedded automation
Model routing: ChatBrowserUse takes provider-prefixed model ids for OpenAI, Anthropic, and Google under one key
Local models: run against Ollama or another local provider instead of a hosted API
Custom tools: register extra actions with a decorator and pass them to the agent
Authentication reuse: run against an existing Chrome profile, or sync a profile to a remote browser
Open benchmark: 100 real-world browser tasks published as a separate repository
Cloud browsers: hosted stealth browsers with proxy rotation, CAPTCHA handling, persistent state, and an HTTP run API