MyClawn

An AI agent that fills out forms

Supplier onboarding, vendor portals, applications, registrations, the same twelve fields on twenty sites — forms are most of the tedious work on the web, and most of where agents quietly fail. Here is what filling one reliably actually takes.

A form looks like the easiest job you could give a machine: fields, values, a submit button. In practice it is where a lot of agents fall over, because a real form is not a list of fields. It is a multi-step, stateful, often logged-in web application with validation, conditional sections, file uploads, a CAPTCHA at the end, and consequences on the other side of the button.

The difference between an agent that can "fill out forms" in a demo and one you would let submit a vendor application is almost entirely about the four things below.

What filling a form reliably requires

1. A real browser, driven like a person drives it

Modern forms are built with the assumption that a person is in a browser: dropdowns that load their options on click, date pickers, address fields that autocomplete, sections that only appear after an earlier answer, uploads that need a real file dialog. An agent that posts a payload to the form's endpoint skips all of that and usually gets rejected or, worse, accepted with half the data. An agent that operates a genuine browser — clicking, typing, waiting for the next section — fills the form the way it was designed to be filled.

2. The data, already on hand

Most of the time in a form is not typing; it is finding what to type. Your VAT number, the bank details, the certificate PDF, the address in the exact format this site wants. An agent with a persistent machine keeps that material on its own desk — the documents you dropped on it last month are still there — instead of asking you for the same things every time. That is the difference between "fill in this form" being a thirty-second instruction and a twenty-minute conversation.

3. A human for the fields a machine should never guess

Some fields are not data-entry; they are decisions or consent. The signature. The "I agree" on terms you have not read. The field where the honest answer depends on judgment. A good agent treats those as stop points: it fills everything it can, then hands you the keyboard on its own screen for the parts that are yours, and carries on after you hand it back. The same handover covers the CAPTCHA and the login — the moments a site is specifically checking for a person.

4. Watching the submit

The submit button is where the cost of a mistake lives. The right design is that you can see the completed form on the agent's screen before it presses the button — and press it yourself if the form matters. On MyClawn the agent's desktop is streamed live to your browser and you can take over at any moment; for the first few forms of any kind, that is exactly what you do. You can watch a recorded run to see the shape of it.

Three kinds of form, three different answers

What still breaks

Where this is the wrong tool

A one-off form you will fill in ninety seconds is faster to do than to explain. A form you fill every week, across many sites, from the same set of documents, is the case — and it gets better the longer the agent's machine keeps your material and your logins. Read how sessions and the take-over work in Security & trust.

Questions people ask

Will it submit the form on its own?

Only if you tell it to. The sensible pattern for anything that carries your name is “fill everything, stop at submit, show me” — you check it in the live screen and press the button yourself.

Does it work on portals with no API?

Yes — that is the point. It fills the same web form you would, field by field, in a real browser, so back-offices and government portals that only exist as pages are in reach.

What if a field is ambiguous?

It asks, or it stops. And because you can see the screen, a wrong guess is caught in the second it happens, not after the form is in.