Best Tools · Browsers
agent-browser 0.39.2
SummaryVercel Labs' browser CLI for AI agents stopped telling agents a click worked when it never reached the page.
Filed October 24, 2026. Price, as published by the maker: Free, open source (Apache-2.0). Runs locally with Chrome; no account or API key needed unless you use its AI chat mode or a cloud browser..
Open the maker’s site (opens in a new tab)
agent-browser is a command-line tool from Vercel Labs that lets AI coding agents drive a real browser: open a page, read it as a short list of numbered elements, then click, type and check results by those numbers. It is a native Rust binary installed from npm, works with any agent that can run shell commands, and has more than 44,000 GitHub stars. Version 0.39.2 shipped October 22, 2026.
How we tested. October 24, 2026, on a Linux box with Chrome 156. We installed 0.39.2 and the previous 0.39.1 side by side and ran the same commands against small local test pages: buttons that were disabled, hidden or placed off the page, a redirect to a blocked domain, and broken policy files. No AI model and no API key. The full test is in the story.
Pro: 0.39.1 said "✓ Done" (and "success": true in JSON) for clicks on a disabled button, two hidden buttons and one placed off the page, though none of them was clicked. 0.39.2 failed all four with a plain reason ("is disabled, so the click would not reach it"), and still clicked the two real buttons. It also now fails open when a redirect lands on a blocked domain, and a broken policy file now blocks commands instead of being ignored.
Con: The docs' own "restrictive" policy example blocked every command on both versions, including close, because it doesn't allow the launch and close actions. We had to kill the background process. After a blocked redirect, get url still shows the blocked address, so check the exit code.