ToolVitals scores Browser Use 94 overall and 100 on shipping, and the public site no longer reads like a single-purpose browser wrapper. It now looks like a platform, with Browser Harness, Stealth Browsers, Browser Use Box, Web Agents, Custom Models, cloud docs, open-source docs, and a benchmark claim of 97% accuracy.
The bet is bigger than automation
The homepage says it plainly, “The Way AI uses the web.” The README frames the choice as open source versus cloud, with cloud browsers pitched for stealth, proxy rotation, and scaling, while the open-source library stays available for custom tools and self-hosting.
That is the real signal here. Browser Use is not trying to be one more helper around Playwright. It is trying to own the runtime where agents touch websites.
Recent releases show the same direction
Release 0.12.6 is full of reliability work, not vanity features. It fixes Gemini-3 defaults, Bedrock structured output, MCP schema compatibility, Windows daemon spawning, cloud reconnect behavior, and an O(n²) DOM capture bottleneck.
Release 0.12.5 is even sharper. It removed litellm from core dependencies after the March 24 supply chain attack on versions 1.82.7 and 1.82.8. That is a clean sign that the team treats dependency risk as part of the product, not as an afterthought.
BUX and Browser Harness explain the roadmap
BUX turns Browser Use Box into a 24/7 remote VM with Claude Code and Browser Harness pre-installed, controlled from Telegram, the web, or SSH. The post says Telegram talks to one resumed Claude Code session, browser sessions rotate every 240 minutes, and the box is reachable through a browser terminal or SSH.
The Browser Harness post pushes the thesis further. It argues for direct CDP access, a thin harness, and letting the model edit its own helpers when something is missing. That is a hard bet on agent autonomy, and it is the opposite of hiding browser internals behind a fat abstraction layer.
What the data does not prove
ToolVitals sees stars, releases, uptime, and public positioning. It does not see code quality, user satisfaction, revenue, or whether Browser Use actually improves task success in production.
It also cannot tell you whether the cloud path is the right tradeoff for your team, or whether the benchmark numbers hold up on your own workflows.
How it stacks up
Compared with LangChain’s 135,175 stars and 29 release events in 30 days, Browser Use is smaller and more specialized. Compared with OpenClaw’s 365,410 stars and 39 release events in 30 days, it is a narrower bet with less surface area.
Browser Use still shipped 17 releases in 90 days and 6 release events in 30 days, so this is not a sleepy repo. It is an active one with a very specific thesis, browsers are where agents live.
Recommendation
If your team needs browser automation for agent workflows, evaluate Browser Use now. If you only need simple scraping or a thin script around one site, this is probably too much platform. If you need remote browsers, stealth, and a path from open source to hosted agents, it belongs on the shortlist.