Published Updated
In one or two sentences, before anything else.
Browser Use is an MIT-licensed Python library and a paid cloud service that let AI agents operate real web browsers, made by a Y Combinator startup founded by Magnus Müller and Gregor Žunič. Its GitHub repository had about 118,300 stars on October 17, 2026, up from about 50,000 at its 17 million USD seed round in March 2025.
What it can do, one sentence each, and whether we actually tested it.
A Python package (Python 3.11 or later) that turns a plain-language task into clicks, typing and navigation in a real browser, using OpenAI, Anthropic or Google models, Browser Use's own model or a local one through Ollama.
MIT license; you pay your own model costs.
Browser Use's own model for browser tasks (bu-2-0), at 0.60 USD per million input tokens and 3.50 USD per million output tokens including Browser Use's fee.
Per Browser Use's pricing page; we found no independent test of the model.
A hosted agent API that runs tasks in Browser Use's browsers, billed at model cost plus 20% plus browser time and traffic.
With your own model key, Browser Use still charges 20% of published token rates.
Remote browsers your own agent can connect to, at 0.02 USD an hour, with residential proxies at 5 USD per GB and CAPTCHA solving.
Browser Use itself says no browser setup guarantees every CAPTCHA can be avoided or solved.
A command-line tool, plus Browser Harness, which connects coding agents such as Claude Code, Codex or Cursor to a real Chrome browser.
Per Browser Use's README and home page.
Agents can pay at a shop checkout with a one-time virtual card from Stripe Link that you approve in the Link app each time, and can pay for Browser Use Cloud itself in USDC over Coinbase's x402, with no signup.
Launched August 6 (x402) and September 3, 2026 (Link); Browser Use says the agent never sees your real card number.
Each cloud agent runs in its own micro-VM with no AWS keys or database credentials inside; real credentials stay in a separate control plane.
Browser Use's post of February 25, 2026. Its enterprise offer lists SOC 2 Type II, HIPAA support and zero data retention, which we did not check.
The people it helps and the problems it solves, and where it ranks.
Built for
Problems it solves
Graded in
Public tests other people ran: what they gave it and how it did.
Case 1
Input
300 tasks on 136 live websites (Online-Mind2Web), each run graded by human evaluators; Browser Use ran on GPT-4o.
What happened
Browser Use completed 30.0% of the tasks, far below the 89.1% its makers had reported on the older WebVoyager benchmark with the same model. Failed runs took nearly twice as many steps as successful ones. This was an early-2025 version, and we found no independent human-judged test of the current one.
Case 2
Input
All 300 Online-Mind2Web tasks with the bu-max cloud agent, graded by an agentic judge Browser Use built on the Claude Agent SDK.
What happened
Browser Use reports 97%, ahead of the 93% OpenAI reported for GPT-5.4 computer use. It is the maker's number with the maker's judge; Browser Use's August 2026 post shows the same runs scoring anywhere from 62.5% to 83.7% depending on the judge.
Case 3
Input
Three simple steps (search, open a page, read a link) on 71 sites behind Cloudflare, PerimeterX, DataDome, Akamai and reCAPTCHA, with the same agent and model on every provider.
What happened
Browser Use Cloud got through 81% of the time, against 77% for Anchor, 67% for Kernel, 47% for Steel, 42% for Browserbase and 40% for Hyperbrowser. The judge only checked whether the agent was blocked, not whether the task was done right, and the benchmark is the maker's own.
Case 4
Input
Let the agent buy an item; at the checkout it asks for payment, naming the amount, merchant and item.
What happened
Stripe Link issues a one-time virtual card only after you approve in the Link app, and the card is discarded after one purchase. This is the vendor's description; we found no independent test. Anthropic's and OpenAI's computer-use guides both advise human confirmation before purchases.
What it does badly, with the evidence, and what we could not settle.
Weak
Independent results are far below the maker's.
Evidence: Ohio State and UC Berkeley researchers counted 30.0% success under human judging on Online-Mind2Web (2025 version, GPT-4o); Browser Use reports 97% on the same benchmark with its own judge (March 2026).
Web content can hijack the agent.
Evidence: A white-box analysis by Mudryi, Chaklosh and Wójcik (arXiv, May 19, 2025) showed prompt injection, domain validation bypass and credential exfiltration in Browser Use, with a working proof of concept.
A critical flaw has already been found.
Evidence: CVE-2025-47241 (CVSS 9.3, published May 4, 2025) let a decoy domain in a URL's username part bypass allowed_domains and reach unauthorised sites, including localhost; fixed in version 0.1.45.
The maker writes, runs and judges most of the benchmarks.
Evidence: Browser Use's own August 2026 post found that eleven judges grading the same 104 runs gave scores from 62.5% to 83.7%, and named the conflict of interest in running models on its own cloud.
Costs swing widely by model.
Evidence: Browser Use's pricing page records agent costs from about 0.06 USD a task (GPT-5.6 Luna) to more than 8 USD (GPT-6 Astra) on its 60-task benchmark, before browser time and proxy traffic.
Non-developers should not use it, and nobody should give it unsupervised access to email, banking or payment accounts.
Evidence: It needs Python 3.11 or later and model API keys. Anthropic's computer-use guide advises a dedicated VM, no access to login details and human confirmation before payments or accepting terms.
Uncertain
Funding since the seed round.
Why unsure: We found only the 17 million USD seed round of March 2025; none of the sources we checked reports a later round.
Usage and customer claims.
Why unsure: The 8.1 million monthly downloads and a logo wall that includes Amazon, Google and OpenAI appear on Browser Use's home page with no customer count or independent figure.
Enterprise compliance.
Why unsure: Browser Use lists SOC 2 Type II, HIPAA support and zero data retention for enterprise customers; we did not see the reports.
Whether stealth browsing is allowed on a given site.
Why unsure: Browser Use sells residential proxies and CAPTCHA solving to get past bot protection; the pages we read do not discuss websites' terms of use, which differ by site.
The numbers, each with its source and the date we checked it.
| Spec | Value | Note and source |
|---|---|---|
| Maker | Browser Use, started in 2024 by Magnus Müller and Gregor Žunič through ETH Zurich's Student Project House; Y Combinator Winter 2025 | Source: techcrunch.comchecked |
| Funding | 17 million USD seed (March 2025), led by Felicis, with Paul Graham, A Capital and Nexus Venture Partners | Source: techcrunch.comchecked |
| License | MIT | Source: github.comchecked |
| GitHub | About 118,300 stars and 13,100 forks | Source: github.comchecked |
| Latest version | 0.13.11 on PyPI (October 7, 2026); requires Python 3.11 or later | Source: pypi.orgchecked |
| Cloud agents | Model cost plus 20%, plus browser time and traffic; no subscriptions | Source: browser-use.comchecked |
| Cloud browsers and proxies | 0.02 USD an hour, billed per minute with a 1-minute minimum; residential proxy 5 USD per GB | Source: browser-use.comchecked |
| Free credit and top ups | 1 USD once per person, no card; 10 concurrent sessions to start; top-ups from 5 USD that never expire | Source: browser-use.comchecked |
| Security advisory | CVE-2025-47241, critical (CVSS 9.3): allowed_domains bypass, fixed in 0.1.45 (May 4, 2025) | Source: github.comchecked |
| Maker's benchmark claim | 97% on Online-Mind2Web with the bu-max agent, graded by Browser Use's own judge (March 25, 2026) | Source: browser-use.comchecked |
Answer first, then why, an example, the limits and what to do next.
The Browser Use library is free and open source under the MIT license. You still pay your model provider for every step; the hosted cloud is pay as you go, with a one-time 1 USD credit, browsers at 0.02 USD an hour and agents at model cost plus 20%.
Browser Use is made by a startup of the same name, started in 2024 by Magnus Müller and Gregor Žunič at ETH Zurich. It was in Y Combinator's Winter 2025 batch and raised a 17 million USD seed round led by Felicis in March 2025, TechCrunch reported.
Browser Use is not safe to point at real accounts without a sandbox and human checks. Researchers showed in May 2025 that web pages can hijack it through prompt injection and leak credentials, and a critical allowlist bypass (CVE-2025-47241) was fixed in version 0.1.45. Run it in an isolated browser, keep logins out of reach and approve payments yourself.
Independent and maker results for Browser Use differ widely. Human judges at Ohio State and UC Berkeley counted 30.0% success on Online-Mind2Web for a 2025 version running GPT-4o; Browser Use reported 97% on the same benchmark in March 2026 with its own judge and cloud agent.
No independent test has compared Browser Use, Stagehand and Playwright MCP on the same tasks. Browser Use hands a whole task to an agent; Browserbase's Stagehand (MIT) adds act, extract and observe steps to code you write; Microsoft's Playwright MCP (Apache-2.0) gives any MCP client browser tools and says it is not a security boundary.
Choose Browser Use for open-source code that works with any model, and the OpenAI or Anthropic tools if you already build on one of those models. Both labs' tools return mouse and keyboard actions that your own code carries out, and both warn that page content can steer the model. Browser Use's March 2026 table puts its agent ahead of OpenAI's, but that is the maker's own comparison.
This page is built from what can be checked today: Browser Use's GitHub repository, home page, pricing page and engineering blog, TechCrunch's reporting, a human-judged study by Ohio State University and UC Berkeley researchers, a security paper and advisory, and the documentation of Stagehand, Playwright MCP, OpenAI and Anthropic. Sources were checked on October 17, 2026.
| Date | Event | Source |
|---|---|---|
| 2024 | Magnus Müller and Gregor Žunič start the company through ETH Zurich's Student Project House | TechCrunch |
| December 15, 2024 | Browser Use reports 89.1% on WebVoyager with GPT-4o, after removing 55 of 586 tasks | Browser Use |
| March 3 to 10, 2025 | Manus, which uses Browser Use, goes viral; daily downloads rise from about 5,000 to 28,000 | TechCrunch |
| March 23, 2025 | 17 million USD seed round led by Felicis; the company is in Y Combinator's Winter 2025 batch | TechCrunch |
| May 4, 2025 | Critical advisory CVE-2025-47241 published, fixed in version 0.1.45 | GitHub |
| September 30, 2025 | Stealth browser infrastructure launches | Browser Use |
| April 9, 2026 | Free tier for remote browsers | Browser Use |
| August 6 and September 3, 2026 | Agents can pay for Browser Use Cloud in USDC over Coinbase's x402, then pay at shop checkouts with one-time Stripe Link cards | Browser Use, Browser Use |
| October 17, 2026 | About 118,300 GitHub stars; version 0.13.11 on PyPI | GitHub, PyPI |
The open-source library is free under the MIT license and needs Python 3.11 or later. You connect your own model (OpenAI, Anthropic, Google, Browser Use's BU 2.0 or a local one through Ollama) and pay that provider for every step. The cloud is pay as you go, with no subscriptions (pricing):
| Item | Price |
|---|---|
| Cloud agents | Model cost plus 20%, plus browser time and traffic |
| Your own model key | 20% of published token rates, plus browser time and traffic |
| Cloud browsers | 0.02 USD an hour, billed per minute |
| Residential proxy | 5 USD per GB |
| Free credit | 1 USD once per person, no card |
| Top-ups | From 5 USD; credits never expire |
Concurrency starts at 10 sessions and rises with total spend, up to 1,000 sessions after 25,000 USD. On Browser Use's own 60-task benchmark, the recorded agent cost runs from about 0.06 USD a task (GPT-5.6 Luna) to more than 8 USD (GPT-6 Astra), before browser and proxy charges.
Every headline number on Browser Use's site is its own. Its December 2024 report claimed 89.1% on WebVoyager with GPT-4o after dropping 55 tasks and correcting the evaluator by hand. Researchers at Ohio State University and UC Berkeley then built Online-Mind2Web, 300 tasks on 136 live websites, and had humans grade the runs (arXiv): Browser Use with the same GPT-4o completed 30.0%.
In March 2026 Browser Use reported 97% on that same benchmark with its bu-max cloud agent, graded by a judge it built itself. Its own August 2026 post shows why that matters: eleven judges grading the same 104 runs gave scores from 62.5% to 83.7%. Browser Use's stealth benchmark (March 21, 2026) had its cloud getting past bot protection on 81% of 71 sites, against 42% for Browserbase; the judge only checked whether the agent was blocked.
| Tool | Maker | License | How it works |
|---|---|---|---|
| Browser Use | Browser Use | MIT; cloud is paid | An agent takes a whole task and drives the browser, with any model |
| Stagehand | Browserbase | MIT | act, extract and observe steps inside code you write; runs on local Chrome, Browserbase optional |
| Playwright MCP | Microsoft | Apache-2.0 | Gives any MCP client browser tools from the accessibility tree, no vision model; "not a security boundary" |
| OpenAI computer use | OpenAI | Paid API | The model returns mouse and keyboard actions that your code carries out |
| Anthropic computer use | Anthropic | Paid API | Screenshots plus mouse and keyboard control, run in your environment |
In Browser Use's March 2026 table, Stagehand scored 55% to 65% and OpenAI Operator 61%; these are the maker's numbers, not an independent test.
Sources
Changelog
Runable is a general AI agent that builds websites, apps, slides, documents, images and videos and now also runs ads and outreach, made by a 15-person startup in Bengaluru founded in 2025 by Umesh Kumar and Saksham Sarda. It turned on paid plans in March 2026 and raised a 21 million USD Series A on August 26, 2026.
Polsia is an AI service from Polsia, Inc. that says its agents plan, code and market a company with no human staff. Founder Ben Broca launched it in December 2025 and reports 10,000 paying customers; critics doubt his revenue figures, and his own numbers show nine in ten companies on it have yet to earn a dollar.
WorkBuddy is Tencent's AI agent for office work, built by the Tencent Cloud CodeBuddy team and launched in China on March 9, 2026. In seven months it added a global edition and phone apps, and Tencent announced a Guangdong government version. It plans tasks, works on your local files and hands back documents, spreadsheets and slides.
Raccoon Work (商汤小浣熊) is SenseTime's AI assistant for office work: data analysis, documents and slides on web, desktop and phone. SenseTime merged its office and coding assistants into it on August 25, 2026, and its first-half results report cloud monthly users up nearly tenfold. No independent test of the new version is public yet.
Muse is Meta's personal AI agent: it books, shops, sends email and fills in forms from its own cloud computer, asking before it pays or sends. Meta launched it in the US on September 8, 2026 and has since added Canada. Early reviews are mixed, and Amazon blocks it from buying on its store.
Manus is a general AI agent that researches and builds slides, websites, videos and games, made by a company founded in China in 2022 and based in Singapore. China blocked Meta's purchase of Manus on April 27, 2026; Manus resumed independent operations on September 1 and launched Manus 2.0 on September 28, 2026.