Browser Use

Browser Use is a cloud platform that turns large language models into fully hosted web agents — software that can navigate websites, click buttons, fill forms, extract data, and complete multi-step workflows inside real browsers. Unlike self-hosted browser automation frameworks, Browser Use operates the infrastructure: stealth browsers, session management, API gateways, and a security-hardened execution environment called AgentCore.

How It Works

Developers interact with Browser Use through a REST API with endpoints for tasks (/api/tasks), browsers (/api/browsers), and sessions (/api/sessions). An agent receives a natural-language instruction, opens a browser session, and executes the task using the same Chrome DevTools Protocol (CDP) that human developers use to debug web applications. The agent writes Python that imports CDP helpers; if a needed helper does not exist, the agent writes it at runtime — the same pattern a developer uses when fixing a missing import.

Security Architecture

AgentCore, the platform's execution environment, isolates each agent in a VM within a private VPC. The sandbox has no direct access to AWS services, databases, or the outside internet. Every external request — LLM calls, file uploads, S3 operations — must pass through a stateless FastAPI control plane. The control plane validates session tokens, swaps them for real upstream credentials, meters usage, and forwards requests. For file operations, the sandbox requests presigned URLs scoped to the session, never holding long-lived AWS credentials.

Benchmark Performance

Browser Use publishes BU Bench, a benchmark that tests web agents on multi-step navigation, search, information extraction, form filling, dynamic UI interactions, iframes, PDFs, and downloaded files across live websites. Anthropic's Claude Fable 5 achieved 80.0% on BU Bench V1 using the open-source Browser Use library, completing 80 of 100 tasks at an average of 6 minutes 53 seconds per task — though the API cost totaled $580.87. The remaining 20 tasks failed due to incorrect fact extraction, insufficient source access, or timeout at the 30-minute mark.

QA Automation

One of the platform's packaged skills is QA automation. Instead of requiring a human to manually test a vibe-coded web application — clicking through every page, noticing what feels wrong, re-prompting the coding agent — the QA skill gives an AI agent a browser, a rubric, and instructions to use the app and report what needs improvement. This closes the loop between code generation and quality assurance.

Learning and Privacy

Agents can learn from past tasks. When an agent discovers a website-specific interaction pattern — like waiting for Google Flights' autocomplete dropdown before pressing enter — it can persist that knowledge as a skill. Before saving, every skill passes through a dedicated PII gate: a separate LLM rejects anything containing emails, tokens, or user-specific data. The platform is also building HTTP-level skills that reverse-engineer underlying APIs from observed traffic, allowing future agents to skip the UI entirely.

For teams evaluating web agent platforms, alternatives in the AI Agents Directory include OPC Directory and PhantomCrew, each with different security and workflow trade-offs.