Browser Use

Browser Use Agents · API V4

Send a sentence. Get the answer back.

One request runs a browser agent on a stealth cloud browser and returns text or JSON. 98% on Online-Mind2Web, the standard live-web benchmark.

Get an API keyAgent quickstart
from browser_use_sdk.v4 import BrowserUse with BrowserUse() as client:    run = client.runs.create(        "Find the cheapest nonstop flight SFO to NYC next Friday. Return airline, departure time and price."    )    result = client.runs.wait_for_completion(run.id)    print(result.result)
main.pyREST · Python SDK · TypeScript SDK
98%
on all 300 live Online-Mind2Web tasks
$0.02 / hour
browser time inside a run, no subscription
1 sentence
is all a run needs to start

A run, start to finish

  1. 1

    Send the task

    Add an output schema if you want JSON, a profile id to start logged in, files, or a proxy country.
  2. 2

    Watch it work

    Stream events for a progress UI, open the live view, or block on completion. A person can take the wheel through the live view for a payment or an approval.
  3. 3

    Get the answer

    Text, or JSON validated against your schema. Follow up in the same session; files land in the workspace. The same browsers are for rent directly through Browser Infrastructure.

BU Bench V2, the hardest public browser benchmark.

Browser Use scores 98% on Online-Mind2Web, the standard live-web benchmark, and saturates the other public task sets. So it built a harder one: BU Bench V2, 200 long multi-step tasks scored against weighted rubrics, the hardest public browser benchmark. Scores are low for every model by design. Current scores on a 60-task cut, running the open-source Browser Use agent: GPT-6 Astra 77.3%, Claude Opus 5 and GPT-5.6 Sol about 50%, GPT-5.6 Luna about 40% at under $0.50 per task. Tasks are encrypted so they stay out of crawlers and training data. Open-source benchmark repository · Every agent benchmark
BU Bench V2: mean rubric score by model against cost per task on a log scale. GPT-6 Astra 77.3%; Claude Opus 5 and GPT-5.6 Sol about 50%; GPT-5.6 Luna about 40%.
BU Bench V2, 60-task cut. Mean rubric score by model against cost per task. From the browser-use repository.

Questions people ask

What do I send, and what comes back?

A task in plain language. Optionally a JSON Schema for the output, files, a saved browser profile, and browser settings such as proxy country. Back comes a run id, ordered events while it works, a live view URL, and the result as text or as JSON matching your schema.

How accurate is it?

98% on all 300 live Online-Mind2Web tasks; the next published result is ABP + Opus 4.6 at 86%. On BU Bench V2, the much harder open-source benchmark on this page, every model scores low by design.

What does a run cost?

Model tokens at the listed rates plus browser time at $0.02 per hour. No subscription. New signups get $15 of credit once.

Can it log in, and get past 2FA?

Yes. Pass a profile id to start logged in, give the run its own inbox with agentmail: true for emailed codes, or bind a 1Password item so passwords and TOTP codes are typed without the model seeing them.

How many runs at once?

10 on a new account, rising with lifetime spend to 1,000, more on enterprise agreements. Browsers and agent runs share one project limit.

Can I follow up?

Yes. A session keeps the browser and the conversation between runs, so the next task continues where the last one stopped. Workspaces keep input and output files.

MCP, n8n, webhooks?

All three. The hosted MCP server exposes runs as tools for Claude Code and Cursor, the n8n node returns typed JSON to the next step, and webhooks deliver events to your backend.