dalao.ai

Browser Use

Make websites accessible for AI agentsOpen source

About Browser Use

Browser Use is an open-source Python library that lets a language model drive a web browser the way a person would: reading the page, clicking, filling in forms, and turning a plain-English task into a sequence of web actions. The library is MIT-licensed and free, and works with OpenAI, Anthropic, Google and other hosted models as well as local models through Ollama. The company behind it came out of ETH Zurich, went through Y Combinator's Winter 2025 batch and raised a $17M seed round in March 2025; it makes money from hosted cloud browsers and an agent API.

EditorialUpdated 11 min read

What it does

Language models are good with words and bad with websites. Plenty of work only happens in a browser: booking travel, checking government lists, downloading invoices, filling in back-office forms that have no API. The traditional answer is a Selenium or Playwright script that clicks through selectors one by one and breaks whenever the site changes. Browser Use lets the model read the page and decide the next step itself.

It runs as a loop. The library turns the current page into something a model can read: a screenshot plus a numbered list of interactive elements, a trimmed version of the DOM and accessibility tree. The model reads the task and the page state and returns one action, such as click element 12, type some text, scroll or extract content. The library carries it out, hands the new page back, and repeats until the task is done or a step limit is reached. The model card for the company's own bu-2-0 model describes exactly this loop.

The browser layer was rebuilt in August 2025. The official changelog (2025-08-19) announced that every interaction now goes straight through the Chrome DevTools Protocol (CDP) with no Playwright dependency, shipped in 0.6.0. Today pyproject.toml lists the company's own cdp-use as the core dependency, alongside SDKs for OpenAI, Anthropic, Google GenAI, Groq and Ollama, plus MCP. There's no Playwright and no LangChain.

Who's behind it: Magnus Müller (CEO) and Gregor Žunič (CTO), who both studied data science at ETH Zurich, according to their YC company page. TechCrunch (2025-03-23) reported that the project began in 2024 at ETH's Student Project House, where the pair built a demo in five weeks and then open-sourced it. Manus, from Butterfly Effect, put it on the map in early 2025, and more than 20 companies in the W25 batch were using it. The company then raised a $17M seed led by Felicis, with Paul Graham, A Capital and Nexus Venture Partners participating.

The product now comes in three layers. At the bottom is the MIT library. In the middle are command-line tools for coding agents: the Browser Use CLI and a separate repo, Browser Harness. On top is the paid Browser Use Cloud, which the official pricing page splits into Browser Use Agents (hand over a task, get the result back; currently API V4) and Browser Infrastructure (managed stealth browsers the customer connects to over the SDK, REST or CDP). Version 0.13.0 (2026-06-09) also added a beta agent rewritten in Rust (from browser_use.beta import Agent); the release notes stress that the existing Python agent is unchanged.

Timeline: the repo was created on 2024-11-01, and the first GitHub release, 0.1.0, came out 2024-11-07. As of 2026-10-26, the release GitHub marks as Latest is 0.13.13, published at 07:50 (UTC+8) on 2026-10-16.

The open-source browser agent.

Browser Use README

Key features

An agent in a few lines. Install it, create Agent(task=..., llm=...) and await agent.run(). Adapters for OpenAI, Anthropic, Google, Groq, Ollama and others are built in, alongside the company's own ChatBrowserUse. The README recommends bu-2-0, which the model card says was released on 2026-01-27 as a mixture-of-experts model built only for browser decisions, not general chat.

Straight to CDP. Since August 2025 browser control has run entirely on the Chrome DevTools Protocol. The company's blog post (2025-08-20) explains that Playwright adds a second hop through a Node.js server's WebSocket, which adds up when an agent makes thousands of CDP calls per step to check element position, opacity, event listeners and the like. Going native made element extraction, screenshots and default actions faster and added proper cross-origin iframe support.

Cheap replays. agent.save_history() stores a successful run, and load_and_rerun() replays it step by step without calling the model. The README warns that saved page snapshots and results may still contain sensitive data. With the sensitive_data parameter, passwords and other secrets appear in prompts only as placeholders.

CLI and coding agents. Version 0.12.3 (2026-03-24) shipped Browser Use CLI 2.0, which talks CDP through a background daemon. Its release notes claim about 50 ms per command and "2x faster, 50% fewer tokens" (vendor-reported, same release notes), and it supports Chrome only. Browser Harness, a separate repo with 18,594 stars as of 2026-10-25, is a CLI that gives AI agents control of the user's own browser. The library can also run as a local MCP server.

Keeping up with model vendors' browser tools. 0.13.11 (2026-10-07) added a toolset for Claude, which the README says covers 31 browser actions, though the release notes admit the anthropic.tools.browser module it needs isn't in the public Anthropic SDK yet. 0.13.13 added an OpenAI browser-tool driver with 18 default tools and 6 optional ones; it needs the openai-tools extra and OpenAI Python SDK 3.31 or later. Both depend on upstream SDKs being ready.

Cloud browsers. Browser Use Cloud browsers come with residential proxies, CAPTCHA handling and profile sync (cookies only, per the README). Each session exposes a live view for human takeover, can record an mp4 and can last up to 240 minutes. The pricing page says the service has passed a SOC 2 Type II audit, and enterprise projects can enable a HIPAA mode.

How it compares

Star counts come from the GitHub API, as of 2026-10-25.

Stagehand (Browserbase): MIT, 25,803 stars, with TypeScript, Python and other SDKs. It hands developers three primitives, observe, act and extract, so the predictable steps stay in code and only the parts that need page understanding go to the model. Browser Use's Agent takes the whole task and lets the model decide each step.

Skyvern: AGPL-3.0, 23,266 stars. It also combines LLMs and computer vision to work websites, offers a Playwright-compatible SDK and a no-code workflow builder for non-technical users, and runs its own cloud.

Playwright MCP (Microsoft): Apache-2.0, 38,592 stars. It isn't an agent; it exposes browser actions as MCP tools for the model inside Claude, Cursor and other clients. Browser Use's own Browser Harness overlaps with it. In a matched evaluation in the Browser Harness README, Browser Harness scored 60.4, Chrome DevTools MCP 61.0 and Playwright MCP 60.0, with no significant difference, and Browser Harness used the fewest tokens (vendor-reported, Browser Harness README).

Hosted browser vendors: Browserbase, Steel, Browserless, Hyperbrowser, Anchor, Kernel and others sell the same kind of cloud browser as Browser Use Cloud. In Browser Use's Stealth Bench V2 (2026-09-25), two visits to each of 100 bot-protected sites got through 90.5% of the time for Browser Use, versus Anchor 82%, Kernel 80.5%, Browserbase 78.5%, Browserless 77.5%, Steel 74.5%, Hyperbrowser 73.5% and headless Chrome 3% (vendor-reported, browser-use.com/benchmarks/browsers). The same page admits the 100 sites were chosen because Browser Use Cloud historically got through where stock Chrome didn't, so they're not a balanced sample. The company's cloud-browser comparison page also notes that Browserless and Steel can be self-hosted, while Browser Use Cloud can't.

Online-Mind2Web: the company reports 98% using its own judge model, against a top human-graded leaderboard score of 97.7%, and says itself that the two used different judges and can't be compared directly (vendor-reported, browser-use.com/best-web-agents).

Who uses it, and who it's for

Adoption (as of 2026-10-25): the GitHub repo has 119,061 stars and 13,228 forks, and the GitHub API's open_issues_count, which covers open issues and pull requests, is 606. On PyPI, pypistats shows 8,061,974 downloads in the past month and 1,329,646 in the past week, counted in UTC through 2026-10-24. These download counts include CI runs and mirrors, so they aren't user counts.

Usage figures for the hosted service come only from the company. Its web-agent comparison page says API V4 handled 1.6 million runs between 2026-09-17 and 10-16, and 95.2% ran to completion. On the 13,807 runs where customers turned on an optional model judge, the judge rated 64% successful, and the company notes this isn't a success rate for all runs (vendor-reported, browser-use.com/best-web-agents). TechCrunch reported that Manus used it for browser automation, and Browser Use's case study says Hermes Agent uses it by default when no browser backend is set (the company's claim).

Who it's for

Python developers building agent products. For agents that work on sites without an API, such as screening government lists, pulling invoices or filling in admin forms, Browser Use provides perception, actions and the loop out of the box, and switching models means changing one llm argument.

Teams that need lots of browsers in the cloud. Browser Use Cloud offers managed browsers with proxies, CAPTCHA handling and anti-detection; Playwright, Puppeteer or any CDP client can attach, so there is no browser fleet to run.

People who work with Claude Code, Codex, Cursor and similar coding agents. The README includes a setup prompt for the agent; together with browser-use skill install, it lets the agent drive the Chrome the user is already signed into.

Anyone staying fully local. The library runs with a local browser and a local Ollama model and never touches a paid service, though results depend on the hardware and model.

Not for

Fixed workflows that only need reliable scraping. Every step calls a model, so it's slower and costlier than a hard-coded Playwright script, and results aren't fully reproducible. The README offers saved-history replay without model calls, but when the page changes the model still has to step in.

Teams wary of data leaving the building who keep the defaults. The docs say telemetry is on by default and may include task instructions, visited URLs, action traces and final results; only ANONYMIZED_TELEMETRY=false turns it off.

Teams that want to host the cloud browsers themselves. The library runs locally, but Browser Use Cloud can't be self-hosted; the company's cloud-browser comparison page says only enterprise customers can discuss on-prem deployment with sales.

Anyone counting on getting past every CAPTCHA. The README says plainly that no browser configuration can guarantee it.

Getting started

Hosted cloud: Browser Use Cloud accepts sign-up with a Google, GitHub or Microsoft account, new accounts get a one-time $1 credit, and tasks can be submitted from the web app or the API.

The library needs Python 3.11 or later. Following the model card: uv pip install browser-use (or uv add browser-use inside a project), then uvx browser-use install to set up the browser runtime, with the model's API key exported. An Agent takes a task and an llm and runs with await agent.run(). Cloud browsers are enabled by passing Browser(use_cloud=True) to the Agent.

Coding agents get a browser through the README's setup prompt: given to Claude Code or a similar tool, it installs or upgrades browser-use, runs browser-use skill install to register the skill, and connects to the user's browser. Troubleshooting steps are in install.md in the Browser Harness repo.

The Rust beta agent installs with uv add "browser-use[core]" and is imported as Agent from browser_use.beta (0.13.0 release notes). The Rust agent is still in beta; the stable agent is the Python one.

Two production-related settings: ANONYMIZED_TELEMETRY=false turns telemetry off, and allowed_domains limits where the agent can go; CVE-2025-47241 was fixed in 0.1.45.

Pricing & license

The library is MIT-licensed (copyright Gregor Zunic, 2024) and free. Model calls cost extra, except with a local model through Ollama.

Browser Use Cloud is pay-as-you-go. Per the official pricing page (as of 2026-10-25): there are no subscriptions and credits never expire. New accounts get a one-time $1 credit, and the free tier allows 10 concurrent sessions with cost-efficient models only. Browsers cost $0.02 per browser-hour, billed per minute with a one-minute minimum. Residential proxies are $5 per GB; direct or bring-your-own-proxy traffic is $0.20 per GB. The BU 2.0 model API costs $0.60 per million input tokens, $0.06 cached and $3.50 output, and a customer-supplied model key adds a 0.2x orchestration fee. Lifetime payments of $100, $1,000, $5,000 and $25,000 raise the concurrency limit to 50, 250, 500 and 1,000 sessions; enterprise terms are custom. The legacy API V2 starts at $0.006 per step plus $0.01 per task.

The sources disagree on one point. The pricing page names gpt-5.6-luna as API V4's default model ($0.24 per million input tokens, $1.44 output), while the API V4 OpenAPI spec gives gpt-6-luna (GPT-6 Luna) as the model field's default. The pricing page is dated September 3; the OpenAPI spec carries only the version number 4.0.0 and no date. The pricing model has changed too: YC's launch post in early 2025 sold the hosted version as a $30-a-month subscription, and it's now purely usage-based, though legacy subscriptions can still be managed.

Risks & limits

First, telemetry is on by default. The variable is called ANONYMIZED_TELEMETRY, but the docs say collected data may include task instructions, visited URLs, action traces, errors and final results, and may be used to build de-identified datasets. Setting ANONYMIZED_TELEMETRY=false turns it off.

Second, security boundaries sit with the user. Advisory GHSA-x39x-9qw5-ghrf (CVE-2025-47241), published 2025-05-04, showed that putting a decoy domain in the username part of a URL bypassed the allowed_domains allowlist and could reach localhost or internal services. It affected 0.1.44 and earlier and was fixed in 0.1.45. GitHub rates it Critical with a CVSS 3.1 score of 9.3. The score shown on NVD comes from the CVE numbering authority, MITRE, at 4.0 (Medium); NVD hasn't scored it itself and lists it as Deferred. The two vectors differ. GitHub calls attack complexity low, with high confidentiality impact, no integrity impact and low availability impact. MITRE calls attack complexity high, with no confidentiality impact, low integrity impact and no availability impact. Another risk is prompt injection: the agent reads arbitrary web content.

Third, it moves fast and the API changes. The engine moved from Playwright to CDP, a Rust beta arrived, and the Claude and OpenAI toolsets depend on upstream SDKs.

Fourth, most performance figures are self-reported. Stealth Bench V2 uses sites the vendor selected, Online-Mind2Web was scored by the vendor's own judge, and the 64% production success figure covers only runs with the judge turned on.

Fifth, compliance sits with the user. Anti-detection browsers, residential proxies and CAPTCHA handling get past many sites' bot defenses, but whether that breaks a site's terms of service or local law is the user's call. The terms of service restrict use in some regions and for some purposes; see the terms for details.

Our verdict

Browser Use is a library for building agents that get things done on the web; it is MIT-licensed and the model can be swapped. The company's paid products are cloud browsers and hosted agents; the library was still shipping releases as of 2026-10-16 (0.13.13), and the marketing numbers are vendor benchmarks. In the default configuration telemetry is on and there is no domain allowlist; versions move fast, and success rate and cost vary by task.

Why it stands out

Browser Use answers a blunt question: how does a language model click on the right thing? The library turns a live DOM into a compact, numbered description of interactive elements, so the model reasons over intent instead of pixels or raw HTML.

Magnus Müller and Gregor Žunič met while studying at ETH Zurich. According to TechCrunch they built the first demo in five weeks and then open-sourced it, and within months it was widely used by agent builders. Since 0.6.0 (August 2025) it has dropped Playwright and drives the browser directly over the Chrome DevTools Protocol (CDP); the vision fallback and the structured action space are all visible in the source.

It gives a model a numbered list of a page's interactive elements instead of pixels or raw HTML.

Sources

Official

  1. browser-use/browser-use (GitHub)GitHub · Accessed Oct 25, 2026
  2. Browser Use READMEGitHub · Accessed Oct 25, 2026
  3. Browser Use LICENSE (MIT)GitHub · Accessed Oct 25, 2026
  4. browser-use/pyproject.toml at main · browser-use/browser-useGitHub · Accessed Oct 25, 2026
  5. Release 0.13.13: OpenAI browser tools and Cloud fixes · browser-use/browser-useGitHub
  6. Release 0.13.11 · browser-use/browser-useGitHub
  7. Release 0.13.0 - Rebuilt in Rust [beta] · browser-use/browser-useGitHub
  8. Release 0.12.3 - Browser Use CLI 2.0 · browser-use/browser-useGitHub

Updates & reading

Newsodysseysbench.com

Browser Use and Carnegie Mellon release Odysseys2, a long-horizon web agent benchmark where the top agent fully completes 47 of 100 tasks

Odysseys2, from Carnegie Mellon and Browser Use, is a web agent benchmark of 100 tasks drawn from Browser Use's real user requests, run on the live web with 982 rubric checks. An LLM judge grades the files agents produce along with their actions and screenshots, and gives zero credit to reward hacking. Of 15 agents tested, GPT-6 Astra leads, fully completing 47 of 100 tasks.

1 source · 1 post on X