Automate web forms with an AI form-filling API

1. Describe the task

   Write the goal the way you would brief a careful assistant: every value the form needs, and what proves success ("tell me the confirmation number"). Start on the form itself when you know its URL.

2. The agent fills the form

   Each step, the agent reads the page and does one action: click, type, choose an option, upload a file or fill a saved secret. Before it clicks a button that submits a form, it checks every field against your goal and fixes values that don't match.

3. You approve, it submits

   By default the agent pauses before purchases, bookings, messages, deletions and applications, and shows you the form's values. You reply with an approval, and the run ends with an answer, plus data in your schema if you asked for one.

curl https://lovelyrouter.com/v1/browser/runs \
  -H "Authorization: Bearer $LOVELY_API_KEY" \
  -H "content-type: application/json" \
  -d '{
    "goal": "Request a quote for 500 blue 750 ml water bottles with our logo, delivered to 12 Main St, Springfield, IL 62701 by December 1, 2026. Contact: Ada Lovelace, ada@example.com. Submit the form and tell me the quote reference.",
    "start_url": "https://supplier.example.com/quote",
    "approve": "submit",
    "output_schema": {"quote_reference": "string", "confirmation_message": "string"},
    "webhook": "https://api.example.com/hooks/browser"
  }'

The request returns 202 at once and the agent works in the background. With "approve": "submit" it stops before the submit button, and your webhook gets a signed run.waiting event whose needs holds a prompt for a person, the form's values in fields, the button label in action and a screenshot. Approve it:

curl https://lovelyrouter.com/v1/browser/runs/brun_5b7d9f1a3c5e7a9b1d3f5a7c/reply \
  -H "Authorization: Bearer $LOVELY_API_KEY" \
  -H "content-type: application/json" \
  -d '{"approve": true}'

The agent clicks the button and finishes; run.finished arrives with the answer and the JSON in data. Reply {"approve": false, "reason": "Wrong address"} instead and the agent doesn't submit: it reads your reason and tries another way or reports why.

Store each credential once with PUT /v1/browser/vault/{name}: its value, a kind (password, username, text or totp) and the origins where it may be typed. Pass the names in the run's secrets. The model only sees names like portal_password; the value is typed only on a site you allowed, so a lookalike login page can't get it.

To watch a run, add "stream": true, "screen": true and get each step and the browser's screen as it happens. When a page shows a CAPTCHA, the agent hands you a Live View link; you solve it, reply {"done": true}, and it carries on. Use the agent only on accounts you own or are allowed to use; the Acceptable Use Policy has the details.

You pay per step: $0.006 in fast mode, $0.012 in accurate mode for date pickers, maps and custom widgets. The models are included and there is no run fee. A login takes 4 to 8 steps; a long multi-page form takes 15 to 25. Each step includes one minute; waiting longer for a person costs $0.005 per minute, and pauses themselves are not billed as steps. A run holds max_steps times the step price plus one wait up to human_timeout_minutes, then settles to what it used. A run that fails or that you stop pays for the steps it took. See pricing, and the browser agent page for the benchmark.