Agent Browser is a free tool from Vercel Labs that lets Claude open web pages, click, type and check its own work in a real browser. I installed it and tested it against Playwright. It is worth it, if you use it the right way and keep your passwords away from the AI.
snapshot -i. The full snapshot is only about a third smaller.Same page (pypi.org/project/requests), November 14 2026, Agent Browser 0.43.0 vs the Playwright MCP snapshot. Characters sent back to the AI, lower is cheaper.
A second page (pypi.org/project/playwright) gave 24,487 vs 15,234 vs 5,690. Playwright MCP also loads 25 tool descriptions (about 21,000 characters) into every chat, which a command-line tool skips.
Needs Node.js 22 or newer. Paste this one line into Claude Code and let it run the setup for you.
Install Agent Browser for me: run "npm install -g agent-browser", then "agent-browser install", then "npx skills add vercel-labs/agent-browser", and confirm it works by opening example.com and taking a snapshot.
Prefer to do it yourself? These are the same steps.
# in your terminal npm install -g agent-browser agent-browser install # downloads Chrome for Testing, first time only npx skills add vercel-labs/agent-browser
Or as a Claude Code plugin: /plugin marketplace add vercel-labs/agent-browser, then /plugin install agent-browser@agent-browser.
Tell Claude to read pages with the interactive snapshot. Add this line to your project's CLAUDE.md so it sticks.
When you use agent-browser, read pages with "agent-browser snapshot -i" (buttons and fields only). Take a full snapshot only when you need the page text.
Agent Browser can log in to sites for you. Never paste a password into the chat. Save it in the built-in vault instead, and the AI only ever sees the name.
# you run this once, in your own terminal (not through Claude) echo "your-password" | agent-browser auth save mysite --url https://mysite.com/login --username you@mysite.com --password-stdin # then Claude only needs this agent-browser auth login mysite
Set AGENT_BROWSER_ENCRYPTION_KEY before saving so the vault is encrypted on disk.
Then tell it exactly which websites it may visit. Anything else is blocked (I tested it: a jump to another site fails with "not in the allowed domains list").
# one run agent-browser --allowed-domains "mysite.com,*.mysite.com" open https://mysite.com # or for every run in this terminal export AGENT_BROWSER_ALLOWED_DOMAINS="mysite.com,*.mysite.com"
The allowlist does not work together with connecting to your own running Chrome (--auto-connect) or a saved Chrome profile. For anything that matters, use the vault plus the allowlist.
Start the live dashboard and open localhost:4848. You see the browser and every command Claude sends, as it happens.
agent-browser dashboard start
Swap mysite for your own site, pick the one job you repeat every week, and save the prompt that works in CLAUDE.md. Next time it is one sentence.
One email unlocks the whole library - every guide, free, forever.
FREE · NO SPAM · YOU'LL ALSO GET OCCASIONAL AI TIPS FROM @AI.WITHMB
Something went wrong - email mb@thewickfirm.com and I'll send you the guide directly.