¶ How it works
Four moments between your agent and a person.
Most of the time your agent should drive the browser by itself. Proxy Human is for the minutes it shouldn't — the captcha, the credit card, the judgment call. Here's exactly what happens.
§ 01 · The wall
Your agent reaches a step it can't take alone.
It starts small — the checkbox the page's anti-automation script refuses to accept programmatically, the date picker that only opens on a real mouse-down. Then it gets harder: the OTP that only lives in your inbox, the Cloudflare Turnstile that's learned to distrust this user agent, the 3-D Secure prompt that wants a fingerprint on the device that owns the card. And then the genuinely subjective — which of these three apartments matches what you said last week, which line item on the receipt to dispute, what wording to use in the appeal letter the agent's about to send on your behalf. Driving blindly past these is brittle; retrying them is slow.
§ 02 · The hand-off
One tool call mints a short, signed link.
The agent calls open_browser_handoff_link. Proxy Human returns a short URL — typically a few-character hmnpr.xyz/… — that you (or a teammate) open. The agent then forwards that link wherever you live: SMS, Slack, Discord, whatever its harness has wired up.
§ 03 · You take the wheel
The browser opens on your phone, exactly as the agent sees it.
A live WebRTC stream of the agent's Chrome tab, plus a way to tap, scroll, type, and navigate inside that same browser session. Cookies, login state, partially-filled form — all of it is the agent's. You're just doing the bits an automated driver can't.
§ 04 · The hand-back
One tap returns control, with a structured log.
Press "return control to agent" and the session closes. Your agent's blocking call returns with the final page URL plus a chronological record of what you did — typed text, presses, taps, navigations — so it can resume with full context, not guesswork.
¶ Ready