BrowserSkill is a command-line tool and browser extension that lets AI agents drive your real, already-logged-in Chrome or Edge browser instead of a fresh automated one. It is aimed at developers who already run coding agents — Cursor, Claude Code, Codex and similar — and want those agents to reach pages that sit behind a login, without minting separate test accounts and without the agent grabbing the tab in front of them. The project comes from Tencent, is written in TypeScript, carries an MIT license, and has collected a little over 4,000 stars since it appeared in mid-2026.
What it does
The pitch is narrow and easy to check: your agent gets browser access, and your own browsing stays undisturbed. Four things follow from that.
- Real login state. Because the automation runs inside the browser you already use, agents work against sites you are signed into. No separate test credentials, no copying cookies around.
- A separate Agent Window. Browser tasks open in their own visible window rather than seizing the tab you are reading. If the agent genuinely needs a tab you already have open, it has to borrow that tab explicitly and hand it back when the task ends, leaving the rest of your browser alone.
- Any agent that can run a shell. The whole surface is the
bskCLI. There is no lock-in to one model, one agent framework or one harness; the README names Cursor, Claude Code, Codex, OpenClaw, CodeBuddy, WorkBuddy, Pi, Hermes Agent and DeepSeek Harness among the clients it connects. - Human-in-the-loop by design. When a run hits a captcha, a login screen, a confirmation dialog or any other human-only step, the agent can stop and ask you to take over, then continue where it left off. The demo video leans on exactly this moment.
There is also a full-page screenshot feature, reachable from the extension's Quick actions menu or from the CLI as bsk screenshot --session <id> --full-page --out page.png, with its own documentation page covering page support, cancellation and export.
How it works
Two local pieces make up the runtime. The first is bsk, which is both the command-line interface and a background daemon. The second is a browser extension loaded into your everyday browser profile, which is why the sessions, cookies and signed-in state are already there — nothing is re-authenticated for the agent's benefit.
The agent side of the contract is deliberately boring. Instead of a protocol adapter or a per-editor plugin, an agent simply shells out to bsk and passes a session identifier, the way the screenshot example above does. Any tool that can execute a command can therefore drive the browser, which is the reason the project can claim such a long list of supported agents without maintaining an integration for each one. In practice the flow shown in the video is: open a real tab that is already logged in, then scrape, click and fill forms in it the way a person would, and escalate to the human when a page refuses to cooperate.
Getting started
Expect to install both halves — the CLI and the extension — and then point your agent at the bsk command. The supported runtimes are listed plainly: macOS on Apple Silicon and Intel, Linux on x64 and ARM64, and Windows on x64. Chrome and Microsoft Edge are the supported browsers; other Chromium-based browsers are expected to work where they can load unpacked Chromium extensions. Firefox and Safari are not in that list. The repository README, which is published in English alongside a Chinese translation, carries the current install commands, and a separate guide covers the long-screenshot feature in detail.
When to use it / when not
This is the right shape of tool when the work is on the authenticated web: checking a staging dashboard you are signed into, pulling numbers out of an internal admin panel, filling a long form, or letting an agent verify a change in the running app rather than in a test double. It also suits people who simply refuse to let an agent take over their working browser, since the separate window is the whole point.
It is a weaker fit for unattended server-side automation. The design assumes a desktop machine with a visible window, an extension installed in a human's browser profile, and a human nearby to clear captchas — none of which describes a headless CI runner. And the same property that makes it useful is the one to think about before installing: an agent holding your live session inherits your real permissions on every site you are logged into, so the blast radius of a bad instruction is your actual accounts, not a sandbox.
If you are building or using shell-capable AI agents and keep hitting the login wall, BrowserSkill is worth a serious look: it solves a specific, common annoyance with a small, legible design — one CLI, one extension, an explicit borrow-and-return rule for your tabs, and a clean handoff when a page needs a human. It is young, and the security trade-off deserves a deliberate decision rather than a reflex install, but for developers who want agents working on the real, signed-in web, this is one of the more honest approaches to the problem.