What you can do here

This is a live, public AI agent built on Human Browser. Type a goal in the prompt box above — the agent runs an isolated stealth Chromium session in our cloud, navigates real websites for you, and streams the browser back to your tab in real time. To run a task you sign in with your email (a one-click link, no password) and add $1 by card — about 4 typical tasks (the median task cost $0.24). After that you top up from $5 and pay only for what you use; there is no free tier.

Available models

The AI that drives the agent is included in the $0.02 step price:

• gpt-6-luna — the default; fast and cheap, used for almost every task.
• Other OpenAI, Anthropic, Kimi and MiniMax models — pin one per task when you need it. A step whose model cost goes above $0.02 is billed at its token cost +30% (see pricing).

Proxy countries

Each session can be routed through residential IPs in different regions. Useful for geo-restricted content, locale-specific pricing, or simply to avoid datacenter detection.

• 75 countries, all residential — the Americas, all of Europe including the CIS and Caucasus (Russia, Ukraine, Belarus, Kazakhstan, Moldova, Georgia, Armenia, Azerbaijan, Uzbekistan, Kyrgyzstan), the Middle East, Asia-Pacific and Africa.
• Romania (RO) — the default: cheapest pool, fast in EU, works for most tasks.
• Switch mid-session — switch_proxy_country relaunches the browser on a new sticky residential IP in ~5s and keeps the profile.
• Direct connection — datacenter IP from the agent host (no proxy).

Pass country=<iso2> (lowercase) at spawn time. Call list_countries for the live catalogue with per-country pool health.

A2A — agent-to-agent integration

Plug in your own AI agent (Claude, GPT, custom LangChain/CrewAI, etc.) via the standard A2A protocol. Send a goal over POST /a2a with a message/send JSON-RPC payload, get back a streamed result and the live browser URL. Works with Claude Agent SDK, CrewAI, LangChain, and raw HTTP clients. See the A2A docs for curl examples andtasks/get resumption.

Built-in challenge handling

The agent automatically detects and handles reCAPTCHA v2/v3, hCaptcha, and Cloudflare Turnstile challenges that appear during navigation — no extra configuration needed. Cloudflare interstitials and DataDome / PerimeterX fingerprint walls are handled by the underlying stealth browser engine.

How it works under the hood

Each goal spins up a fresh Chromium instance running our patched stealth build (real desktop Chrome fingerprint, residential IP, real timezone + WebGL hardware). The chosen LLM drives a browser-use loop: it inspects the DOM, picks the next action, and the orchestrator executes it via Playwright. The browser screen is captured by ffmpeg + xvfb x11grab and streamed to your tab as multipart MJPEG so you can watch the agent work step by step.