browser-use: Detailed Introduction

[ikaijua]

browser-use: Detailed Introduction

This post is maintained as the product evolves. The browser-use entry in the repo README links here, so detailed info is updated here instead of churning the repo.

browser-use lets AI agents operate a real web browser through natural-language instructions. Agents can click, type, fill forms, navigate across pages, and extract data.

• Open-source library: a Python library under the MIT license, available at browser-use/browser-use.

• Cloud service: Browser Use Cloud provides hosted agents and stealth browsers.

• Related tools: Browser Harness, Box, CLI, and MCP.

• Founders: Magnus Müller and Gregor Žunič, ETH Zurich classmates who released the project in late 2024.

• Community: over 100,000 GitHub stars in 2026, making it one of the most-starred open-source browser-agent projects.

• Company: Browser Use Inc. joined YC W25 and raised a $17M seed round led by Felicis in March 2025.

Official account & key contributors on X

• Browser Use — official (@browser_use)

• Magnus Müller — cofounder (@mamagnus00)

• Gregor Žunič — cofounder (@gregpr07)

The sections below expand on "what it is → three ways to use → Cloud deep-dive → ecosystem → strengths & limits → quick start". Click the contents to jump to each floor.

Table of Contents

1. What it actually is
2. Three ways to use it
3. Browser Use Cloud
4. Ecosystem
5. Strengths & limitations
6. Quick start
7. Comparisons & community reception

Click the contents to jump to each floor; floor content is maintained as the product evolves.

[ikaijua]

What it actually is

In one line: it "translates" a website into a structured list of elements an agent can understand, then lets any LLM click, fill, read, and operate across pages.

• Core mechanism — DOM distillation: instead of giving the model a screenshot (costly, slow, error-prone), it breaks the page into a compact representation of interactive elements (buttons, inputs, links), sharply cutting token use and improving decision accuracy.

• Model-agnostic: Claude, GPT, Gemini, and local models (Ollama, etc.); the team also ships its own browser-tuned models, ChatBrowserUse / BU2.

• Extensible: register custom tools with a @tool decorator (replacing the older Controller), with lifecycle hooks and tracing.

• Background: open-sourced in late 2024, prototyped in about 5 weeks; joined YC W25 in 2025 and raised a $17M seed (led by Felicis) in March.

[ikaijua]

Three ways to use it

• Open-source Python library — local or self-hosted, free, MIT, runs on your own or a remote browser, any model, with custom tools, hooks, and tracing.

• CLI 2.0 (Mar 2026) — dropped Playwright for direct Chrome DevTools Protocol (CDP); ~2× faster and ~50% fewer tokens, built for coding agents like Claude Code and Codex.

• Cloud API (hosted) — hand a task to a hosted agent that returns results and files, no browser infrastructure to run.

• A MCP server also exposes navigation, form-filling, extraction, and tab management as native tools to Claude Code and similar clients (tutorials cover n8n, Vercel, OpenClaw, Hermes Agent, and more).

[ikaijua]

Browser Use Cloud

Two products, pay-as-you-go since 2026 (no subscription):

① Browser Use Agents (API V4)

• Give a plain-English task; the agent writes code, runs long multi-step, multi-page work, and returns the result and files. V3, by contrast, operated step-by-step on the page.

• Rerunnable scripts: the first run saves the verified steps; when a page changes, rerun with new values instead of calling the model every time.

• Optional LLM judge scores each run.

• Human checkpoints via live view.

• An iMessage channel; choose US or EU (Frankfurt) deployment.

② Browser Infrastructure (managed stealth browser)

• Managed stealth Chromium over CDP, also drivable by Playwright / Puppeteer / Selenium.

• CAPTCHA solving (reCAPTCHA + hCaptcha, off in the EU), residential proxies across 195+ countries, persistent profiles, live view and recording.

• REST API + Python/TS SDK + CLI + MCP.

• In the team's own Stealth Bench V2, across 200 visits to 100 bot-protected sites, headless Chrome passed 3.0% of the time vs 90.5% for Browser Use.

Compliance: SOC 2 Type II, zero-data-retention, EU region, DPA; HIPAA on enterprise.

Pricing

• Browsers: $0.02 per browser-hour, with no subscription.

• Residential proxies: $5 per GB.

• Agents: model token costs plus a 20% service fee.

• Credits: start at $5 and never expire.

• Minimum billing unit: 1 minute.

• Concurrency: 10 sessions initially, scaling to 1,000 with lifetime spend.

• New signups: a one-time free credit, with no card required.

[ikaijua]

Ecosystem

• Browser Harness — a self-healing harness that gives Claude Code / Codex / Cursor your real Chrome (spun out as its own project).

• Browser Use Box (BUX, Apr 2026) — a 24/7 remote VM bundling Claude Code + Browser Harness, reachable via web terminal, SSH, or a Telegram bot.

• Own models — ChatBrowserUse and BU2, tuned for browser automation.

• Rust rewrite in progress — an experimental browser_use.rust Agent aims for memory safety, parallelism, and token efficiency (the current Python build still has gaps in memory use, anti-bot, and parallel management).

• Community & adoption — 1,200+ community automation templates; used by teams at Airbnb, Amazon, and Anthropic; 20+ YC W25 companies run it in core workflows.

• License note — workflow-use is AGPL, unlike the MIT library; mind the difference for commercial use.

[ikaijua]

Strengths & limitations

Good for:

• Developers / Python-capable tinkerers who want maximum flexibility in browser automation;

• Technical teams doing affordable large-scale web scraping and data extraction;

• Any scenario where an AI product needs to interact with websites;

• Lean side-project ops doing competitor research, price comparison, repetitive web tasks.

Be careful if:

• You don't know Python and don't want to learn (the open-source library will block you);

• You need 100% determinism for critical tasks (browser-agent paths aren't fixed, harder to debug);

• A simple need is already solved by Zapier / a browser extension.

Known limitations

• Chrome is memory-hungry and parallel agents are resource-intensive.

• The real world throws CAPTCHAs and anti-bot walls (Cloud's stealth browser solves this but adds cost).

• Runs are non-deterministic, so critical production work needs care.

• Against closed alternatives (OpenAI Operator, Anthropic Computer Use are closed APIs), browser-use offers an open, auditable, modifiable, self-hostable foundation.

[ikaijua]

Quick start

from browser_use import Agent
import asyncio

async def main():
    agent = Agent(
        task="Go to Amazon and find the cheapest USB-C hub with 4+ ports",
        llm=ChatBrowserUse(),   # or Claude / GPT / Gemini / a local model
    )
    result = await agent.run()
    print(result.final_result())

asyncio.run(main())

• Install: pip install browser-use (or uv add browser-use); requires Python 3.11+.

• Docs: https://docs.browser-use.com

• Cloud: https://browser-use.com

• GitHub: https://github.com/browser-use/browser-use

[ikaijua]

7. Tests and comparisons on X

Hermes before and after integration

Teknium (@Teknium, Nous Research cofounder and Hermes Agent developer) | August 11, 2026 (UTC+8)

• Comparison: Hermes's existing browser tools, which execute individual actions, versus a new mode powered by Browser Use CLI 3.0, where the agent writes a script to execute browser operations.

• Results: Teknium reported approximately 60% fewer tokens per task in internal tests, with no accuracy drop. Nous Research's accompanying announcement reported a 48–66% reduction.

• Scope: This compares two ways of organizing browser tools within Hermes. It does not establish that Browser Use is more efficient than Playwright or Stagehand across all tasks.

• Limitations: The posts do not provide the full model configuration, task set, or individual run results. Teknium is part of the development team; these are the integrating team's internal results.

Original X post · Teknium's test explanation · Nous Research announcement

Browser Use's internal model comparison

Browser Use (@browser_use) | October 18, 2026

• Tasks: long-horizon browser tasks.

• Score: the team reported 75/100 for DeepSeek V4.1 Flash 1017, matching Opus 5.5.

• Cost per task: $0.22 for DeepSeek and $2.35 for Opus.

• Scope: this compares models, rather than browser automation frameworks.

Original X post

Cost-ratio discrepancy

• The post says “19x cheaper,” but $2.35 ÷ $0.22 ≈ 10.7.

• @haolv511482 flagged the discrepancy and asked whether the headline used a different run from the chart. The listed costs do not support the 19× claim.

Original X reply