Tencent's open-source BrowserSkill lets an AI agent borrow your own logged-in browser tab for repetitive tasks, instead of routing the work through paid cloud browser infrastructure like Browserbase.
A revenue ops person who needs to pull numbers out of five ad platforms, a CRM, and an internal admin panel every Monday usually has two options: do it by hand, tab by tab, or stand up browser automation infrastructure to do it for them. Tencent's BrowserSkill offers a third path. It lets an AI agent borrow a tab in the browser you already have open, already signed into, run the task, and hand the tab back, instead of routing the work through a paid cloud browser service like Browserbase, whose Developer plan starts at $20 a month and its Startup plan at $99 a month.
The project's own description is plain about what it does: it connects agent harnesses including Cursor, Claude Code, Codex, OpenClaw, CodeBuddy, WorkBuddy, Pi, Hermes Agent, and DeepSeek Harness to a browser you're already logged into, so the agent can act inside sites without a separate test account. The mechanism is a local CLI and daemon called bsk paired with a browser extension. An agent calls bsk through the shell, the daemon routes the request to the extension, and the extension carries out the task in a separate, visible "Agent Window" while your own browser windows stay untouched. When a task hits a captcha, a login screen, or a confirmation dialog, the agent can hand control back to you and continue once you've cleared it.
That local, human-in-the-loop design is also the production-readiness signal worth checking before recommending anything like this. BrowserSkill is MIT licensed, sits at 4,000 stars and 284 forks, and its most recent tagged release, bsk CLI 0.3.0, shipped the same day this was checked, one of 19 releases since the project's initial public release in June 2026. That's an active maintenance cadence, not a weekend demo.
Where this actually saves money is in what it avoids provisioning. A cloud browser platform like Browserbase charges by the browser hour and by concurrency: its free tier includes one browser hour and three concurrent browsers, its $20-a-month Developer plan includes 100 browser hours before metering kicks in at $0.12 an hour, and its $99-a-month Startup plan includes 500 hours before $0.10 an hour. Browserbase itself frames 100 hours as roughly 3,000 page-level tasks, which is useful context for how that meter adds up for a team running a handful of recurring internal jobs rather than a scraping fleet. BrowserSkill sidesteps that meter entirely because the browser doing the work is the one already running on your machine, under your own login.
This fits a specific kind of team, not every team. A marketing ops person automating ad platform reporting, a RevOps person exporting CRM data on a schedule, or a small agency doing repetitive data entry across a handful of client admin panels can install the CLI and extension and get value fast, especially because there's no separate service account to provision for every tool the agent needs to touch. It's a worse fit for a company that wants to run a fleet of scraping or testing agents at real concurrency. Cloud infrastructure exists partly to solve that scaling problem, and BrowserSkill, tied to one browser profile on one machine, isn't built to replace it there.
The honest caveats are worth naming plainly. BrowserSkill currently supports Chrome and Microsoft Edge, with Firefox listed as planned, so a team standardized on another browser is out of luck for now. Self-hosting isn't free in the way "no invoice" implies: someone still has to install the CLI and extension on each machine, keep the daemon running, and in a sandboxed or CI environment do extra setup to keep it alive between commands. The project's own documentation notes that phone-only QR scans, face verification, unavailable SMS codes, and image-only captchas can still block a text-only model, which is a reasonable limit but a real one. And BrowserSkill's materials don't claim the compliance features that show up on Browserbase's higher tiers, things like a HIPAA business associate agreement or SSO, which matters if the workflow touches regulated data.
The interesting part isn't that an agent can click around inside a browser. That's been possible for a while. It's that reusing a person's own authenticated session, rather than spinning up a separate one in the cloud, is now a deliberate design choice with real cost and access tradeoffs on both sides. Worth asking before adopting either approach: who else, besides the person who logged in, now effectively has that login.