Use case
If a person can click it, your agent can too.
Midjourney, your ad manager, the supplier portal, half the tools your team pays for: no API, no export, no way in except the browser. The agent operates them the way a person does, with a map of the page instead of raw HTML and a pointer that moves like a hand.
What the agent can do
One tool per thing a person does with a mouse and a keyboard. The names are the ones the agent really calls.
- Click something
- click · press_and_hold
- Type into a field
- type_text · fill_form
- Use the keyboard
- press_key
- Pick from a dropdown
- interact
- Read the page
- snapshot · read_text · read_value
- Copy and paste
- clipboard
- Take files with you
- download_file · save_pdf
- Wait for something
- wait_for · scroll
A map of what can be operated
The answer carries the buttons, fields and links with their roles and names. The next step carries only what changed, not the page again.
Anything the browser draws next to the page
Permission bubbles, device pickers and download bars get answered instead of stopping the run. A standard automation tool cannot even see them.
The result comes back as a file
Downloads, screenshots and PDFs land in the run, with a still image for every step that produced them.