toolbrowser-agentautomationpythonweb-agentsllmopen-sourceaccessibilitydomUpdated 2026-07-06

Browser Use (Framework)

The dominant open-source framework for giving LLM agents control of a real browser. "Make websites accessible for AI agents." 96.4K GitHub stars as of 2026-05-31, the most-starred browser-agent project.

View original tweet

What it is

Browser Use is a Python library that lets an LLM drive a web browser to complete tasks - navigate, read the page, click, type, and extract data - by exposing the page to the model in an agent-friendly representation rather than raw HTML. The tagline is "make websites accessible for AI agents." A developer wires up a model and a task, and the framework runs the perceive-decide-act loop against a live browser. Its star count (101,568 as of 2026-06-30) makes it the reference implementation most of the ecosystem reaches for first. Source: browser-use/browser-use README, https://github.com/browser-use/browser-use, 2026-06-30

Current source snapshot: latest GitHub release 0.13.2 published 2026-06-12; default branch main at 2454d3e2551705232333c906ded8fc31ab0fc9f2 from 2026-06-28. Source: GitHub browser-use/browser-use, 2026-06-30

How it differs from Kevin's primary tool

Kevin's enforced default for agent browsing is Browser Testing Skills (a native Rust CLI using CDP, token-efficient ref snapshots, React introspection, Web Vitals, batch commands). Browser Use is the popular Python-framework alternative:

  • Language and shape - Browser Use is a Python library embedded in agent code; agent-browser is a CLI any agent or CI can shell out to.
  • Ecosystem position - Browser Use is what much of the open-source agent world standardizes on, so it is the relevant comparison point and integration target, even where agent-browser is preferred for Kevin's workflows.
  • Related products - distinct from Browser Use Box (Bux) (an always-on personal browser agent box from a different author) despite the shared name fragment. The Browser Use org also ships Video Use, a sibling agent skill for editing existing video footage.

For repeatable site playbooks see Autobrowse (Browserbase Skills); for visual-regression agent loops see Browser Harness - Self-Healing Browser Automation. A Browser Use agent can also be a tool node inside an orchestrator like LangGraph.

Browser Use CLI 3.0

The July bookmark promotes Browser Use CLI 3.0 as a coding-agent browser-control surface: a terminal agent can connect to real Chrome or a cloud browser, drive it through the Browser Use stack, and keep the browser session visible enough for human debugging. The local thumbnail shows a split between Chrome and a Claude Code-style terminal, which is the relevant product claim for Kevin: browser operation is becoming a first-class CLI adjunct for coding agents, not only a Python library embedded inside a custom agent. Source: X/@browser_use, 2026-07-02

View original tweet

Current source check on 2026-07-06: browser-use/browser-use default-branch HEAD is 18484f23ac96bb955259a1c54530a7d265dfffdb. The Browser Use browser-harness project remains the closer comparison to Kevin's local Browser Harness - Self-Healing Browser Automation because both emphasize direct browser/session control for coding agents. Local routing stays unchanged: use Kevin's Browser Testing Skills and Browser Harness - Self-Healing Browser Automation first unless the task explicitly asks for Browser Use, Browser Use Cloud, or the CLI's own integration surface. Source: GitHub git ls-remote, 2026-07-06

QA skill pattern

Browser Use's upstream qa skill is an evaluator pattern, not merely another browser-control wrapper: it QA-tests a website or web app and returns a 1-5 quality score with evidence. The skill explicitly warns that a text-only model cannot judge screenshots by itself; no-vision agents should delegate visual judgment to Browser Use v2 cloud agents or another vision-capable reviewer, using pass/fail judging plus structured scoring. Source: Browser Use qa skill docs, 2026-06-30

Install path from upstream:

npx skills add https://github.com/browser-use/browser-use --skill qa

For Kevin's stack, this changes routing narrowly: use Browser Harness - Self-Healing Browser Automation or Browser Testing Skills first for local visual QA unless the user asks for Browser Use, a Browser Use cloud/subagent flow is required, or a text-only builder needs a multimodal reviewer. The reviewed artifact pairs GLM 5.2 as a text-only builder with Browser Use v2 multimodal QA subagents that review the site, find bugs, judge aesthetics, and send targeted feedback. Source: X/@browser_use, 2026-06-20

Timeline

Related Pages

Browser Testing SkillsBrowser Use Box (Bux)Video UseAutobrowse (Browserbase Skills)Browser Harness - Self-Healing Browser AutomationBrowser Automation and QALangGraph