
Browser Use
Browser Use is a browser automation platform for AI agents. Describe a task in plain English and a hosted agent runs it in a managed stealth browser, returning structured JSON. Open-source Python library, CDP-compatible browsers and purpose-built models included.
What is Browser Use?
Browser Use is the infrastructure layer that lets AI agents operate a real web browser. It comes in three families, separated by who controls the browser and where the agent runs. Hosted Agents, reached through API V4, take a goal written in ordinary language, carry it out in a managed stealth browser and hand back a structured result; the caller keeps control of the task, the model, the files, the browser settings and the output schema. Browser Infrastructure drops the agent entirely and rents remote managed Chromium browsers, driven from the Python or TypeScript SDK, a REST API, or over the Chrome DevTools Protocol, so existing Playwright, Puppeteer and Selenium scripts connect without modification. The open-source library, installed with pip install browser-use, is for teams who would rather build and run the agent themselves, with their own tools, prompts, hooks and output schemas. The performance claims are unusually specific and published with their methodology: 98% on Online-Mind2Web across 300 tasks on 136 live sites with none removed, against 86% for the next result; 81% of protected sites reached on an in-house stealth benchmark of 71 sites, against 77%; and 84.8% on the 296-task BrowserBench suite. Stock headless Chrome reaches 2% of those protected sites. Models are interchangeable per run — Claude Opus 5, Grok 4.5, GPT-5.6, Gemini 3.6 Flash and MiniMax M3 are hosted, or you connect your own provider key. Browser Use also trains its own BU 2.0 models for browser work, claiming 68 seconds per task against 225 to 330 for general-purpose computer-use models, and 53 tasks per dollar against 2. Around this sit residential proxies in more than 195 countries, CAPTCHA solving inside the session, persistent profiles, cookie sync, live preview, session recordings, deterministic replay scripts that cost no tokens, and human takeover for logins, approvals and payments. Sub-second cold starts and per-minute billing from USD 0.02 an hour make wide fan-out practical: concurrency runs from three sessions on the free tier to five hundred on Scaleup. The company is SOC 2 Type II certified and publishes a machine-readable llms.txt alongside its documentation.
What it does
- Run a web task described in a single sentence inside a real browser and return the result
- Return schema-validated JSON, with a webhook callback when a run finishes
- Fan out thousands of agents in parallel, each with its own browser, proxy and model context
- Provide remote stealth Chromium browsers driven over CDP, an SDK or REST
- Solve CAPTCHAs inside the session and route traffic through residential proxies in 195+ countries
- Keep persistent browser profiles and hand control back to a human for logins, approvals or payments
- Watch a page continuously and report when it changes, with no scraper to maintain
When to use Browser Use / When not to
A quick filter to help you decide if Browser Use is the right fit.
When to use Browser Use
- Developers building web agents who want an open-source Python library they can run themselves
- Data and platform engineers pulling structured data out of websites with no selectors to maintain
- QA and test automation engineers reusing existing Playwright, Puppeteer or Selenium scripts over CDP
- Teams whose agents must reach anti-bot protected sites at scale, up to 500 concurrent sessions
- Startups, indie builders and researchers who need a permanent free tier and pay-as-you-go usage with no subscription
When not to use Browser Use
- Non-technical users looking for a no-code interface: everything goes through code or an API
- Anyone who needs a mobile app, since no iOS or Android application is offered
- Buyers who require a fixed monthly cost, as billing meters browser minutes, proxy gigabytes, tokens and credits separately
- Organisations that cannot accept model training on submitted inputs, for which no opt-out is documented
- Teams needing a written retention period or a named hosting country outside an Enterprise contract
How to use Browser Use
A typical end-to-end flow, from setup to results.
- Pick the product family according to who should control the browser: hosted agent, managed browsers alone, or the local library
- Create an account on cloud.browser-use.com; the free tier asks for no card
- Generate an API key from the dashboard settings
- Hosted agents: POST to /api/v4/runs with the X-Browser-Use-API-Key header, a plain-language task and a model name
- Declare an output schema so results arrive as valid JSON rather than prose
- Collect the result from the event stream or from the webhook fired when the run ends
- Browser infrastructure: POST to /api/v3/browsers, optionally with a proxy country code, then connect to the returned CDP URL from Playwright or Puppeteer
- Stop sessions explicitly, because browser time is billed by the minute
- Open-source route: pip install browser-use, then build the agent with your own model and browser
- Follow runs through live preview and session recordings, taking over by hand for a login or a payment
Pros & Cons
Pros
- Benchmark results published with open methodology and task lists, not just headline percentages
- The core Python library is open source with 108,500 GitHub stars, so the whole loop can be self-hosted
- Very low entry cost: USD 0.02 per browser-hour, a permanent free tier and no mandatory subscription
- Existing Playwright and Puppeteer scripts connect over CDP without a rewrite
- Best measured stealth of the panel tested: 81% of protected sites reached, against 2% for stock headless Chrome
- Per-run model choice and bring-your-own-key avoid lock-in to a single LLM provider
- SOC 2 Type II with a named auditor and observation window, plus a published DPA with SCCs and a sub-processor list
Cons
- The privacy policy states that inputs are used to train the models, and no opt-out is documented
- No retention period is quantified outside an Enterprise contract
- No hosting country or region is named: the DPA allows processing wherever the company or its sub-processors keep facilities
- No postal address, legal notice page or contact page is published, only email addresses
- Multi-meter billing across browser hours, proxy gigabytes, tokens and credits makes the cost of a real task hard to forecast
- Strictly for a technical audience: no no-code interface and no mobile application
- The terms impose individual arbitration under Californian law and waive class actions
Pricing & Plans
Browser Use offers a permanent free plan covering ten agent tasks per month, free browsers and three concurrent sessions. The cheapest paid tier is Dev at USD 29 per month, which grants USD 29 of monthly credits and twenty-five concurrent sessions. Usage may also be purchased without any subscription, through non-expiring credits, with browser time billed at USD 0.02 per hour.
- 10 agent tasks per month
- free browsers
- 3 concurrent sessions
- purchased non-expiring credits
- 10 concurrent sessions after the first top-up
- no monthly task cap
- USD 29 of credits per month
- 25 concurrent sessions
- advanced stealth
- top-ups and bring your own key
- USD 299 of credits per month
- 200 concurrent sessions
- USD 999 of credits per month
- 500 concurrent sessions
- managed proxy at USD 4/GB
- annual credit pool
- SLAs
- data retention terms
- dedicated support
- via sales@browser-use.com
- browser session USD 0.02/hour billed by the minute
- managed proxy USD 5/GB (USD 4/GB on Scaleup)
- proxyless or bring-your-own-proxy egress USD 0.20/GB
- MiniMax M3 0.36/0.072/1.44
- Gemini 3.6 Flash 1.80/0.18/9.00
- Grok 4.5 2.40/0.36/7.20
- Claude Opus 5 6.00/0.60/30.00
- GPT-5.6 6.00/0.60/36.00 (USD)
- provider cost plus a 0.2x orchestration fee
- USD 0.60/1M input
- USD 0.06/1M cached input
- USD 3.50/1M output
- from USD 0.006 per LLM step plus USD 0.01 per task initialisation
- two months free
- with credits granted up front
Data, GDPR & hosting
A consolidated view of how Browser Use handles your data.
GDPR overview
Browser Use never claims GDPR compliance in so many words, but it publishes a full data processing agreement, last updated 30 July 2026. That DPA names the GDPR and the UK GDPR among the privacy laws it covers, incorporates the European Commission's 2021/914 Standard Contractual Clauses by reference, and handles the UK Addendum and the Swiss FADP explicitly. Sub-processing runs under Clause 9 Option 2, with ten days' notice before any new sub-processor and a five-day right of objection; the current list is published on a Vanta trust page. Data subject rights cover access, correction, deletion, portability, objection, withdrawal of consent, appeal and complaint to a supervisory authority. The company states it neither sells personal information nor shares it for interest-based advertising. No Article 27 EU representative is named, and Do Not Track signals are not honoured. Privacy contact: privacy@browser-use.com.
Who owns the data?
You decide entirely what you submit: the agent filters nothing, much as a browser or a search engine would not. Browser Use acts as controller for its own site and account data, and as processor — or sub-processor — for the customer data it handles under contract, which the privacy policy expressly does not cover. Inputs and outputs may be passed to the third-party LLM providers that power the service, and to hosting, analytics, payment and support vendors. Professional advisers, authorities and a future acquirer may also receive personal information. Browser profiles retain cookies and session state, though the vendor states it does not store browser passwords.
Reuse rights
Outputs come back as structured JSON you can pass straight to the next step, and nothing in the terms requires permission to reuse them. Two limits matter. The terms forbid using the service to harvest email addresses or contact details of other users for unsolicited messaging, and forbid circumventing any block Browser Use applies, including through a proxy or a VPN. You remain solely responsible for the legality, quality and accuracy of what you submit, and Browser Use may refuse or remove content at its sole discretion. Separately, it reserves the right to build aggregated, de-identified, anonymous or sanitised datasets from inputs, outputs, usage telemetry and workflow patterns, and to use, share and license them.
Data retention & training
Hosting summary
No hosting country or region is named anywhere on the site. Section 11(a) of the data processing agreement states that customer personal data may be processed in any country where Browser Use, its affiliates and its authorised sub-processors maintain facilities; for data covered by European privacy law, it undertakes not to transfer outside the country of collection without first putting the necessary measures in place. The European Commission's Standard Contractual Clauses are incorporated by reference, with the UK Addendum and the Swiss FADP handled separately. The current sub-processor list is published on a Vanta trust page cited in section 7(a). Stripe processes payments, and inputs and outputs may be passed to third-party LLM providers. The responsible entity is a US company whose team sits in San Francisco, with Californian law governing the terms. SOC 2 Type II attests to the controls without specifying a storage jurisdiction. The domain resolves to an anycast address in the United States, which identifies a delivery node rather than a place of storage.
Things to keep in mind
Risks and trade-offs to weigh before adopting Browser Use.
- Inputs feed model training according to the privacy policy, with no documented opt-out: anything the agent reads or types can end up improving the product
- The agent filters nothing, so it is easy to hand it personal or sensitive data, including other people's, without meaning to
- Persistent profiles and cookie sync mean a logged-in session left in place exposes a real account to an unintended automated action
- Agents can fill forms and place orders, so a poorly bounded run can trigger something irreversible before a human notices
- Stealth, anti-detection and CAPTCHA solving carry legal and contractual risk against the target sites' own terms, and invite retaliatory blocking
- Usage billing across several meters means an agent stuck in a loop can spend with no obvious ceiling
- A 98% benchmark score invites over-trust: the remaining 2% pass unnoticed without human verification
Setup & Integrations
Technical difficulty
Straightforward for a developer, and impossible without one. The open-source route is a single command, pip install browser-use. On the cloud side, an account, an API key and one HTTP POST are enough to launch an agent or a browser. Existing Playwright, Puppeteer and Selenium scripts connect over CDP without a rewrite. There is no no-code interface. Expect to manage session lifecycle, model choice, output schemas, credits and concurrency yourself, and to keep the Cloud SDK and the open-source package apart, since the documentation warns that their APIs differ.
Deployment
Integrations
Supported languages
Behind Browser Use
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Browser Use.
Frequently asked questions
What exactly is a web agent?
Is there a free plan, and what does the cheapest paid tier cost?
How is browser time priced?
Which models can the agent use?
Do my existing Playwright or Puppeteer scripts still work?
Is Browser Use open source?
Is my data used to train the models?
How long is data retained?
What security assurances exist?
Is there a minimum age, or a mobile app?
Should you pick Browser Use?
Browser Use has moved well beyond the open-source library that made its name. It now sells three things at once — hosted agents, managed stealth browsers and its own purpose-built models — and it is the rare vendor in this space that publishes benchmark methodology alongside the numbers. For a technical team that needs an agent to act on real, defended websites at scale, the combination is hard to fault: CDP compatibility means existing Playwright and Puppeteer code keeps working, USD 0.02 per browser-hour keeps fan-out affordable, per-run model choice avoids provider lock-in, and the open-source path remains open to anyone who would rather run the whole loop in-house. SOC 2 Type II certification with a named auditor, a published data processing agreement carrying Standard Contractual Clauses, and a public sub-processor list are more than most tools of this age offer. The reservations concern data and predictability rather than capability. The privacy policy says plainly that inputs are used to train the models, and no opt-out is documented; retention is quantified only inside a contract; and no hosting country is ever named. The company publishes no postal address, no legal notice page and no contact page, which is thin for a vendor asking to drive authenticated browser sessions. Billing spans browser minutes, proxy gigabytes, tokens and credits, so forecasting the cost of a real workload takes a pilot rather than a spreadsheet. Recommended for developers, data and QA engineers, and AI teams comfortable with an API-first product. Less suitable for non-technical buyers, anyone needing mobile access, or organisations with firm contractual requirements on retention or data residency who are not ready to negotiate an Enterprise agreement. Watch the pricing page and the changelog: this product moves fast.
- Choosing a selection results in a full page refresh.
- Opens in a new window.