Browser Use

Browser Use ships as three layered products: the open-source browser-use Python library (MIT-licensed; an agent loop that converts a goal into browser actions, for self-hosted use with your own LLM keys), Browser Infrastructure (remote Chromium sessions with stealth, CAPTCHA-solving, proxy and profile features, drivable over CDP via Playwright/Puppeteer/Selenium or the vendor SDK), and Browser Use Agents, a hosted Agent API (V4) that accepts a natural-language task and returns a result without the caller writing any browser-automation code. A hosted MCP server exposes the agent to MCP clients.

• Hosted agents (API V4): natural-language task in, result back, with JSON structured output, multi-turn sessions and a per-session follow-up queue
• Workspace file management and rerunnable scripts for repeated extraction runs
• Human-in-the-loop takeover via the live browser, plus documented 2FA and 1Password handling
• Managed cloud browser infrastructure: remote Chromium controllable via Python/TypeScript SDK or standard Playwright/Puppeteer/Selenium over CDP
• Custom stealth Chromium fork, built-in CAPTCHA solving at no extra charge, and residential proxies in 195+ countries (or bring your own HTTP/SOCKS5 proxy)
• Browser profiles for persisted login state, live preview, and session recordings
• Hosted MCP server (streamable HTTP) for Claude Code, Claude Desktop, Cursor, Windsurf and other MCP clients
• Open-source Python agent loop (browser-use library, MIT) with 15+ LLM providers and customizable tools/prompts
• Webhooks (documented for V2 tasks and V3 sessions) and scheduled V4 agent jobs

Developers building AI agents or automation pipelines that need browser access - ranging from teams that just want a hosted API to run a task, to teams that want raw cloud browser infrastructure under their own automation code, to teams that want to self-host the open-source agent library entirely.

Pricing is entirely usage-metered (browser-hours, proxy GB, and model tokens plus a 20% fee for hosted agents) with no subscription plans offered, so a team that needs a fixed monthly bill has to negotiate Enterprise terms. Managed residential proxy traffic is billed at $5/GB on top of browser-hour and token costs, which stacks up for proxy-heavy scraping workloads compared to flat-fee scraping APIs. Concurrency starts at 10 sessions and only rises with lifetime spend, so high-parallelism jobs need either sustained spend or an Enterprise agreement.

Setup assessment: 5m

The cloud Agent API quickstart is: create an account, generate an API key from the dashboard, export BROWSER_USE_API_KEY, and send a POST /runs request - no OAuth flow. The open-source library and cloud browser infrastructure paths take longer to configure (LLM provider keys, proxy/profile setup) but are not required for the basic hosted-agent flow.

Prerequisites

• Browser Use Cloud account and API key
• For the open-source library: an LLM provider API key (OpenAI, Anthropic, etc.)

Authentication

API key sent via the X-Browser-Use-API-Key header (keys start with bu_, no Bearer prefix)

Base URL

https://api.browser-use.com/api/v4

Rate limits

Concurrency and HTTP rate are separate. Concurrency starts at 10 sessions and rises with lifetime spend (up to 1,000). Per-project HTTP budget for general traffic defaults to max(25, 2x concurrency) RPS capped at 100; edge WAF allows 1000 RPS per source IP. A session queue holds 20 pending messages.

Machine-readable API

OpenAPI specification linked

• POST /runs - create and execute a hosted agent task
• GET /runs/{run_id}/status - lightweight status poll
• GET /runs/{run_id} - retrieve the full run result
• POST /sessions/{session_id}/queue - queue a follow-up message to a session
• POST /browsers - provision a cloud browser session (returns cdpUrl)
• PATCH /browsers/{session_id} - control browser session lifecycle (e.g. stop)
• /workspaces - file upload and workspace management
• /profiles - create and manage browser profiles for preserving login state

• No official YouTube channel could be confirmed (none is linked from the site footer)
• Rate-limit details were taken from the docs; the pricing page and docs disagree on the spend needed for 50 concurrent sessions ($200 on pricing page vs $100 in docs), so the tier ladder beyond the 10-session start is not stated in pricing fields
• The docs MCP page lists 7 tools while the marketing /mcp page says 6 (it omits list_browser_profiles); tool names are taken from the docs page