The cloud browser API that finishes the task, not just opens the page

How it works

1. 1

   Send a goal

   POST a message/send JSON-RPC call to the A2A endpoint with your token. The task is plain English. Optional metadata: country, profile, mobile_ua, model.

2. 2

   Watch it run

   The response carries metadata.session_id and metadata.viewer_url. Open the viewer to see the browser live. If the agent needs an OTP, the task moves to input-required and you can answer or take over.

3. 3

   Get the result

   Poll with tasks/get and wait_seconds, or use message/stream for server-sent events. States: working, input-required, then completed, failed or canceled.

curl -X POST https://agent.humanbrowser.cloud/a2a \
  -H 'Authorization: Bearer hb_live_xxxx' \
  -H 'Content-Type: application/json' \
  -d '{
    "jsonrpc": "2.0", "id": 1, "method": "message/send",
    "params": {
      "message": { "role": "user",
        "parts": [{"kind":"text","text":"Open news.ycombinator.com and return the top 5 story titles as JSON."}] },
      "metadata": { "country": "us" }
    }
  }'

curl -X POST https://agent.humanbrowser.cloud/a2a \
  -H 'Authorization: Bearer hb_live_xxxx' -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":2,"method":"tasks/get","params":{"id":"<task id from result.id>","wait_seconds":30}}'

Or plug it into your agent over MCP

Add the remote MCP server to Claude, Cursor or Codex. Tools: humanbrowser_run (starts a task, returns task_id and viewer_url without blocking), humanbrowser_status (long-polls up to 60 seconds) and humanbrowser_viewer_url. Running locally instead: npx -y @virixlabs/humanbrowser mcp with HB_TOKEN set.

{ "mcpServers": { "humanbrowser": {
    "url": "https://agent.humanbrowser.cloud/mcp",
    "headers": { "Authorization": "Bearer hb_live_..." } } } }

Agent card: agent.humanbrowser.cloud/.well-known/agent-card.json · OpenAPI: /openapi.json · Full reference: /llms-full.txt

What is in one API call

A headless browser API hands you a browser. This one hands you a browser, the agent that drives it, and the parts that usually break a web task in production.

What teams send it

Built for developers and companies shipping AI agents, data extraction, QA and monitoring. Each example below is the literal text you would put in the task.

Pricing

Prepaid balance, no subscription, no monthly minimum. Your first top-up can be $1, later top-ups from $5; balance never expires.

Measured on our production: 132 successful tasks run by 35 customer accounts in September 2026, our own and benchmark traffic excluded, re-priced at the current rates. Cost per task includes failed attempts in the same session. About 13% met a captcha.

Pay by card through Stripe (Apple Pay and Google Pay included) or in crypto (USDT, USDC, BTC, ETH, SOL). Refund terms are on the refund policy page.

When a raw headless browser API is the better fit

Per raw browser-hour we now charge about what Browserbase, Steel and Hyperbrowser charge: $0.10 an hour. Browser Use Cloud is cheaper at $0.02 an hour. The difference is what sits on top: on Human Browser the agent is $0.02 per step with the AI included, and the residential IP and captcha handling are part of the service. If you already run your own Playwright or Puppeteer script and just need a hosted Chrome to connect it to, a raw browser service with no agent costs less. Prices below are from their own pricing pages, as of Sep 2026.

They are also stronger on compliance (SOC 2, and HIPAA and SSO on Browserbase Scale), on concurrency (25 to 100 and up on paid plans, against our 10 per token), and on raw CDP access. Where we come out ahead is residential traffic ($4 per GB against $10–12 on Browserbase and $6–10 on Steel), no monthly fee, and the work you do not have to write: the agent, captcha handling, human takeover and persistent logins on one bill. Steel Launch is cheaper than us on captchas ($3 per 1,000 against our $5).

So compare on cost per finished task and on engineering time, not per hour. Detailed comparisons: Human Browser vs Browserbase, vs Steel, Browser Use alternative, Browser Use Cloud pricing, Skyvern alternative.

Limits

• Concurrency: up to 10 sessions at once per paid token. Need more? Write to [email protected].
• Goal-level API: you send tasks, not browser commands. There is no raw CDP or WebSocket endpoint to connect your own Playwright or Puppeteer script to.
• Files: up to 8 files per task, base64 attachments up to 25 MB.
• Captchas: if automatic handling fails, the task asks for a person in the live viewer rather than guessing.

FAQ

Related