Agent browser

The Assistant and any MCP client can open real browser windows and work through a website the way a person would: read the page, click, type, take screenshots. It's useful for testing a site you built in Dev mode, or for chores on a site with no API.

Turn it on

The browser is off for agents until you tick Let MCP clients and the assistant run these in Settings. That's the same switch the Automation operations use.

Web pages can talk to your agent

Whatever a page says goes back to the model. A page can carry text written to trick an agent, and with this switch on an agent can also run programs. Only leave it on with agents and sites you trust.

See what's running

When an agent has a browser open, a globe and a count appear at the bottom right of the window. Click it to see each browser's name and address.

• The eye shows a hidden window so you can watch it work, and hides it again.
• The stop button stops a page that's still loading.
• The X kills that browser. Kill all kills every one.

Agents open browsers hidden unless they ask otherwise. Up to five can be open at once, and they all close when Kitlig quits.

What a browser can and can't do

Each browser starts empty, with its own cookies and storage, and loses them when it closes. It never sees what you're signed into, in Kitlig or anywhere else.

Only http and https addresses open, so a page can't reach the files on your disk. Local servers work, which is what you want for testing. Downloads are cancelled, and sites asking for your camera, microphone or location are refused.

Screenshots are saved in the project as intermediate files under images/agent-browser, so a screenshot needs a project open.