Ask anyone who's built browser automation what the hardest part is, and it's rarely clicking buttons or filling forms. It's login.
Captchas, SMS codes, QR scans, device fingerprints, risk control on new devices — a whole wall designed to keep machines out. The usual workarounds all hurt: test accounts (trigger even stricter risk control), cookie hijacking (fragile, expires), headless browsers (obvious fingerprints, instantly blocked).
BrowserSkill (1.8k stars, MIT, by Tencent) takes a different path: it doesn't open a new browser. It borrows the browser you already use, already logged in, already trusted by every site you visit.
The idea
BrowserSkill is two local components — a bsk CLI/daemon and a browser extension. Once installed, your existing agents (Cursor, Claude Code, Codex, and even WorkBuddy) can drive your real, logged-in browser through the bsk command.
Three things I care about:
- Reuse real login state — no test accounts. It borrows trust itself: a device with months of normal history and a valid session succeeds far more often than any freshly-minted environment.
- A separate, visible Agent window — browser tasks run without interrupting your own work, and you can see what the agent is doing. No black box.
- Built-in human-in-the-loop — when a task hits a captcha, login, or confirmation dialog, the agent stops and asks you to take over, then continues.
The real distinction
This is the key difference from Playwright/Puppeteer-style headless automation:
- Headless browsers create a new instance — fingerprints, cookies, and login state are all new and suspicious.
- BrowserSkill borrows your existing instance — fingerprints and sessions are old and trusted.
That's not a performance optimization; it's a paradigm shift from faking trust to borrowing trust. The two are complementary: scrape public data with headless, and hand the "must be logged in" steps to BrowserSkill.
The honest caveat
It reuses your real logged-in browser, so security matters. The design is deliberately conservative — an agent must explicitly borrow a tab, return it when done, and leave the rest alone, and it asks you to take over for sensitive steps. But when it's operating on accounts that touch money, privacy, or public publishing, keep an eye on it. It offers visibility and takeover, not full autopilot.
I've localized the README and core docs to Chinese: https://github.com/yangshun2005/BrowserSkill-cn
If you find this project useful, a star on the original repo supports the author's ongoing maintenance.
Top comments (0)