The open-source library that lets AI agents drive the web — 110k+ GitHub stars
HQ: San Francisco, CA (founded via ETH Zurich Student Project House) | Founded: 2024 | Employees: Small seed-stage team | Stage: Seed ($17M led by Felicis, March 2025); YC W25 | Website: https://browser-use.com
Browser Use is an open-source framework that lets AI agents operate web browsers the way humans do — reading pages, clicking buttons, filling forms and completing multi-step tasks. Its core trick is representation: rather than relying on screenshots and computer vision (which are slow, expensive and fragile), Browser Use extracts the structure of a web page — every button, field and link, with stable IDs — and feeds it to the LLM as text. Agents then 'see' the page the way the DOM does, and can re-run tasks reliably at a fraction of the cost.
The project was born at ETH Zurich in 2024, where founders Magnus Müller and Gregor Žunič were doing master's degrees in data science and met through the Student Project House accelerator. Müller had spent years building web-scraping tools; the pair wondered what happened if you combined scraping with LLM-driven agents. The prototype took five days, launched on Hacker News, and the project was open-sourced almost immediately.
Scale arrived through someone else's spotlight: in March 2025, Butterfly Effect's Manus agent — the most-discussed AI product of that season — used Browser Use as its browser layer. Traffic and stars exploded; Y Combinator (W25) took the company into its batch, and a $17M seed closed within weeks, led by Felicis Ventures with Paul Graham, A Capital, Nexus Venture Partners, YC, Liquid2, SV Angel and the Pioneer Fund participating.
Today the project spans an open-source Python/TypeScript library and a commercial layer: hosted remote browsers for agents (from $0.02 per browser-hour) and a hosted agent API for teams that want web automation without managing infrastructure. Real-world uses range from login automation and data extraction to QA testing and CRM workflows. As agents move from demos to daily work, Browser Use is positioning itself as the fundamental interface layer between models and the web.
Browser Use owns the most precious commodity in the agent era: the last mile between a language model and the actual internet. Agents can reason brilliantly, but until they can fill in the form, click the button and survive the redesign, they can't do real work. Browser Use's insight was deceptively simple — don't make agents look at pixels; hand them the webpage's structure as clean text with clickable element IDs. It's cheaper, faster and far more reliable than vision-based navigation, and it made the project the connective tissue of the 2025 agent boom.
The catalyst was accidental fame. In early 2025, Butterfly Effect's Manus — the Chinese agent that became a global sensation overnight — used Browser Use under the hood. Every new Manus user became a Browser Use user; stars rocketed past 50,000 within months and 110,000+ by 2026, making it one of GitHub's most-starred AI projects ever. More than 20 companies in a single YC batch built on it. Paul Graham personally joined the $17M seed led by Felicis' Astasia Myers — who had also led Chroma's seed — alongside A Capital, Nexus, YC and SV Angel.
The business is now two-sided. The open-source library stays free and remains the de-facto standard for web automation in Python and TypeScript. The company sells the infrastructure agents need at scale: a hosted remote-browser cloud at $0.02 per browser-hour and a managed agent API, so teams don't run browser farms themselves. The strategic tension is textbook infra: the web is hostile to robots (LinkedIn and others actively change their sites to break agents, and legal exposure around ToS-violating automation is unresolved), while model vendors keep building native browser use into their own products (OpenAI's Operator, Anthropic's computer use). Browser Use's bet is that an open, cheap, model-agnostic layer — with enterprise features like proxies, CAPTCHA handling and observability — stays the neutral utility of the agent internet, the way stripe of web-driving is worth owning even if every model can sort of drive.
Browse: Home | Blog | About | All Startups A-Z | View full profile on StartupWiki