Pick it if you are a developer or automation team that needs an AI to finish multi-step website work, such as forms, logins and data pulls, without maintaining selectors. It replaces brittle Playwright or Selenium scripts on sites that keep changing. The vendor's own 180-task benchmark scored its agent 72.9 out of 100, so test it on your sites first.
What changed
Browser Use is an open-source browser agent with more than 117,000 GitHub stars that lets a language model operate a real web browser from a plain-English task. It also runs as a hosted service with stealth browsers and returns results as structured JSON. The core library is MIT licensed and written in Python.
Maker: Browser Use · Autonomy: semi autonomous · Maturity: GA
Underlying models: GPT-5.6 Luna (hosted default), Claude Opus 5, Claude Sonnet 5, Grok 4.5, GPT-6 Astra, GPT-5.6, Gemini 3.6 Flash, MiniMax M3, BU 2.0 (Browser Use model API)
About Browser Use
Browser Use started in 2024 as an open-source Python library from Magnus Müller and Gregor Žunič, two founders who built it through ETH Zurich's Student Project House accelerator. The company raised a $17M seed round led by Felicis Ventures in March 2025, and its careers page describes a team of six. It sits between a tool and an agent: the library is free to run yourself, and the hosted Browser Use Agents product sells the same idea as a managed service with browsers, proxies and CAPTCHA handling included.
What makes it an agent rather than a scraper is the loop. You describe the job in plain English, an LLM reads each page, chooses what to click or type next, and keeps going until the task is done, then returns JSON that follows a schema you define. Selectors someone wrote by hand are not involved, so a redesigned page does not break the run. Hosted runs let you choose between models from Anthropic, OpenAI, xAI, Google and MiniMax, including Claude Opus 5 and Claude Sonnet 5, and Browser Use also sells its own web-agent model, BU 2.0, as a direct API.
The best-fit work is anything a person does across several pages with no API behind it: filling forms, signing in to portals, pulling product data, checking a booking site, or running a QA pass. Developers who already have automation code can skip the agent and use only the remote browsers through Playwright, Puppeteer or Selenium. Rivals split the same job differently. Skyvern leans on prebuilt workflows, Browserbase sells hosted browsers for agents you write yourself, Manus aims at open-ended research and documents, and MultiOn is a closed-source agent delivered through a Chrome extension. For context on the open-ended end of the field, read what Manus is for.
It is also becoming plumbing for other agents. According to Browser Use, Hermes Agent ships it as its default browser tool, and OpenClaw agents use it for stealth browsing. Through a separate add-on for coding agents, tools such as Claude Code get a real browser as well. Release 0.13.11 on 7 October 2026 added toolsets for Claude, and in August 2026 the cloud gained scheduled V4 agents and an Agency skill that asks for approval only before external actions such as sending or publishing.
Pricing
No subscription. Browser Use Agents charge model tokens plus a 20% service fee (the hosted token rates on the pricing page already include it). Browser time is $0.02 per hour, billed by the minute with a one-minute minimum, and managed residential proxy traffic is $5 per GB or $0.20 per GB with your own proxy. The minimum top-up is $5 and credits never expire. With your own model key you pay the provider directly and Browser Use adds 20% of the published token rates. Concurrency starts at 10 sessions and rises with lifetime spend: 50 at $200, 250 at $1,000, 500 at $5,000 and 1,000 at $25,000, with custom limits on enterprise agreements. The open-source library is free, and BU 2.0 is sold as a direct model API at $0.60 per million input tokens and $3.50 per million output tokens. Read from the vendor's pricing page on 2026-10-13; the billing page is authoritative for a purchase.
Key Features
- One-sentence task runs: Post a task to API V4 and the agent returns JSON that follows a JSON Schema you supply, with webhook callbacks when a run finishes.
- Stealth browser sessions: Each run gets an isolated session on a hardened Chromium fork with per-session fingerprints and built-in CAPTCHA solving; the vendor reports reaching 81% of 71 protected sites in its own benchmark, against 2% for stock headless Chrome.
- Human takeover and live view: Every session exposes a live URL so a person can watch or take over for logins, approvals and payments, and recordings are kept for audit.
- Secrets kept from the model: Credentials pass as sensitive_data placeholders and 1Password vault logins fill programmatically, so the model sees the placeholder and the browser receives the value.
- Model choice per run: Pick any hosted model for each run, or bring your own provider key on eligible plans and pay the provider directly.
- Scheduled and watching runs: V4 agents can be scheduled with pause and resume controls, which suits a task like checking a page and reporting when it changes.
- Open-source core: The MIT-licensed Python library runs on your own machine with a local or cloud browser, so the agent logic is not tied to the hosted service.
- SOC 2 Type II audit: An independent firm audits the service and the report is available under NDA; HIPAA support, zero data retention and EU data residency are offered on enterprise agreements.
Strengths
- The MIT licence means little lock-in: a team can start on the hosted cloud and move the agent loop in-house later without rewriting its tasks.
- Billing is pay as you go with credits that never expire, which suits spiky workloads because quiet months carry no minimum.
- The benchmark runner and task sets are published on GitHub, so a buyer can rerun the test on their own sites instead of trusting a headline score.
- Agent, browser, proxy and CAPTCHA handling sit under one account and one concurrency limit, which keeps procurement and debugging simpler.
Weaknesses
- Every benchmark on the vendor's site is run by Browser Use itself, and the 82% strict accuracy on its 106-task Internal Bench Hard set had no independent replication in our search.
- Cost per task swings with the model: the vendor's pricing-page run recorded $7.371 for Claude Opus 5 against $0.126 for GPT-6 Luna medium (60 tasks, model cost only), so budget with a pilot, not a list price.
- Residential proxy traffic is on by default and billed per gigabyte, so data-heavy pages can cost more than the browser time.
- Regulated-data needs are gated: a team handling health records or EU residency rules must negotiate an enterprise contract, because the self-serve tier does not list those controls.
Frequently Asked Questions
What does Browser Use actually cost?
There is no subscription. A hosted run costs the chosen model's tokens plus a 20% fee, and the meter also runs on browser time at $0.02 an hour and on proxy traffic at $5 per GB. Starting hosted rates range from $0.24 per million input tokens on the default GPT-5.6 Luna to $6.00 on Claude Opus 5, and enterprise agreements are quoted individually.
Does Browser Use have a free plan?
Yes, in two forms. The open-source Python library is MIT licensed, so you pay only for your own model and browser. On the cloud, eligible Google, GitHub or Microsoft sign-ups get a one-time credit of $1 on the pricing page (the GitHub README advertises a different figure, so check the billing page), limited to cost-efficient models and 10 concurrent sessions; the first top-up unlocks every model.
What should you use instead of Browser Use?
Skyvern suits repeatable portal workflows with stored logins and recordings. Browserbase fits teams that already wrote their agent and need hosted browsers only, and Manus is the pick for open-ended research and documents. Stagehand, a TypeScript library that adds AI actions on top of Playwright, suits developers who want more deterministic control than a free-roaming agent.
What separates Browser Use from Skyvern?
Both are open-source browser agents with a hosted cloud, but they bill differently: Browser Use meters model tokens, browser hours and proxy traffic separately, while Skyvern's HokAI record lists freemium plans with a workflow builder. Browser Use reports 72.9 on its own 180-task bench and Skyvern reports 64.4 on WebBench, yet the test sets differ and both are vendor-run, so the two cannot be ranked against each other. Choose Browser Use to swap models per run and Skyvern for prebuilt portal workflows.
How do you get started with Browser Use?
For the cloud, sign in with Google, GitHub or Microsoft, create an API key and post a task to the V4 runs endpoint; the vendor's quickstart is a single curl command. For the library, run uv add browser-use on Python 3.11 or newer, add a model key to a .env file and hand the agent a task string. The Cloud SDK and the open-source package are separate APIs, so do not mix their code samples.
Top Alternatives
- Skyvern: Pick Skyvern for repeatable portal workflows with stored logins and run recordings; pick Browser Use when you want to choose the model on every run.
- Browserbase: Pick Browserbase if you have already written the agent and just need rented browsers; pick Browser Use if you want the agent loop supplied with them.
- Manus: Pick Manus to hand over a loose research or writing request; pick Browser Use for defined web tasks whose JSON output feeds your own code.
- MultiOn: Pick MultiOn for a closed-source agent that lives in a Chrome extension; pick Browser Use when you want open-source code you can host yourself.