Browser Use

BU Bench V2.1.1 includes all merged updates since v2.1: the default BrowserCode runner, Cloud and local browsers, Luna xhigh, 100-task concurrency, CAPTCHA-aligned rubrics, the dataset cleanup, and the validated Reuters playback redesign. Download **Source code** for the complete runnable repository, or follow the [setup instructions](https://github.com/browser-use/benchmark/blob/v2.1.1/README.md#running-bu-bench-v21-default): ```bash git clone --branch v2.1.1 --single-branch https://github.com/browser-use/benchmark.git cd benchmark ``` After installing the documented dependencies and setting your own API keys, run: ```bash uv run python run_eval.py ``` The default is all 200 tasks, BrowserCode 0.1.20, OpenAI GPT-6 Luna xhigh, Browser Use Cloud, the GPT-5.6 Luna xhigh findings judge, 3,600 seconds per task and up to 100 concurrent tasks. Local Chrome and optional Laminar reporting remain supported. GitHub Actions runs the same entry point. ### Dataset and scoring - Includes the CAPTCHA

BU Bench V2.1 freezes the current 200-task dataset with the merged task/rubric corrections. Use this tag for reproducible comparisons; the compatible dataset filename remains `BU_Bench_V2.enc`. ### Dataset changes - Nine tasks have corrections relative to the original release: 014, 028, 029, 048, 053, 088, 113, 163 and 187. These cover search completeness, source-backed aliases, supporting evidence, task wording, changing source counts and blocked property lookups. - AliExpress 171 and Reuters 185 restore their original rules after the earlier blocked-access edits. Reporting a block does not itself complete an acquisition requirement; Reuters retains its original evidenced-absence outcome. - All 200 task IDs, item IDs, weights, canaries and 55-task subset membership are unchanged. - Instructions changed for 014, 028, 029, 088 and 113. Older executions are not executions of those revised instructions. ### Judge and runner The public runner defaults to findings judging and a 60-minute ta

Learn how Chrome DevTools Protocol (CDP) works, when to use it, and how to control Chrome with a step-by-step WebSocket and Puppeteer example.

Unsure which browser infrastructure to pick? Learn about Browser Use and Browserbase, their main use cases and how they compare in 2026.

Vesta's AI agents use Browser Use to search OFAC, SAM.gov and HUD lists, pull county tax bills and save each page as a PDF on the loan, under zero data retention.

Composio Browser Tool runs every task on a Browser Use agent in a managed cloud browser, so agents connected through Composio can log in, fill forms and download files on websites with no API.

As models get better, the browser harness has to change. A lot.

In this article, we'll look at the most popular browser automation tools. We'll compare their use cases, performance, browser support, and language support.

Material Event filed 2026-09-10

Learn what a headless browser is, how it compares to a headful one, its use cases, and the basics of browser automation.

## Browser Use 0.13.10 - Upgrade Browser Harness to 0.1.13. - Exact-pin all declared runtime, optional, development, and build dependencies. - Upgrade and migrate to MCP Python SDK 2.1.1. - Add pydantic-settings 2.15.0 as an explicit runtime dependency. - Pin Pydantic 2.13.5 and Hatchling 1.32.0. - Upgrade pypdf to 6.16.2, resolving the three current pypdf Dependabot advisories. - Report unknown MCP tool calls as application errors instead of successful results. Validation includes the full GitHub test matrix, model adapters, MCP stdio handshakes, cross-platform wheel installation, focused PDF tests, CodeQL, GitGuardian, and an installed-runtime vulnerability audit.