Once your assistant has a browser, it can do real work on the web. It's worth knowing how it goes about that, because browsing behaves differently from everything else your assistant does.
This page assumes you've already picked a browser. If you haven't, start with Configure a Browser for OpenClaw.
How your assistant uses a browser
Your assistant doesn't know what a page contains until it looks. Every step goes through the same loop:
- Look. It reads the page and works out what's actually on it: the fields, the buttons, the links, the text.
- Decide. It compares that against what you asked for and picks one action.
- Act. It clicks, types, scrolls, or navigates.
- Wait. The page reacts, and often loads more content.
Then it looks again. A task you'd describe in one sentence, like "find me a flight to Lisbon next Friday", is a dozen or more of these loops.
Here's an assistant doing exactly that, searching for flights:
This recording is trimmed. We cut some of the waiting out by hand to keep the demo short, so the real thing runs slower than it looks here.
Why it can be slow
Each of those loops takes real time, and the slow part is usually look, not act.
Reading a page isn't like reading a file. Your assistant has to take in everything the page offers and make sense of it before it can choose a single click. A modern site can carry thousands of elements, most of them irrelevant.
On top of that, pages move. Content loads late, the page changes, banners appear. Your assistant often has to look more than once before the page settles enough to act on.
So expect web tasks to take minutes, not seconds. A search with a few pages of results is normal work for it, but it's doing a lot of looking.
Logging in
Your assistant can log in on its own if you hand it the credentials. It's better if you don't.
Log in yourself instead, in Host mode. Bring up the browser window with Show browser in the chat header, go to the site, and sign in exactly as you normally would. Use your password manager. Complete two-factor if it asks. Your assistant is not reading what you type.
That way you're giving it the environment, not the credentials: a browser that's already signed in, ready to work. It can use the session without ever knowing your password, and your password never goes through a chat message.
The session sticks. It's saved in the assistant's own browser profile, so from then on it's already signed in and can get straight on with the task. You do this once per site, not once per task.
This only works in Host mode.
Bot protection
Some sites try to work out whether a visitor is a person, and treat you differently if they decide you're not. They might block the page, hand you a challenge, or quietly serve a worse version of the site.
Host mode is the best way to avoid it, because what they're checking is largely real. It's a genuine Chrome on a genuine Mac.
Beyond that, what helps is you navigate on that browser profile. Bot protection leans on whether a visitor looks established, and a profile that has been signed in, has cookies, and has been used before looks a lot more established than one opened seconds ago. So:
- Log in to the site yourself, first. A signed-in session is the strongest signal that someone real is here, and it's the one that makes the most difference.
- Clear any challenge by hand. Once you pass, the site usually trusts that session and stops asking.
- Let the profile accumulate. The trust carries over, because the profile persists. A profile you've used a few times gets challenged less than a fresh one.
The pattern is that the first visit to a demanding site is the one that needs you. Sign in, clear whatever it asks, and after that your assistant can usually work on its own.
Captchas
Captchas were built for a web where every visitor was a person clicking for themselves. That's no longer quite the world we're in: your assistant is working on your behalf, on a task you asked for. The site just has no way to tell.
In practice your assistant will struggle to get past one, so the quickest path is to solve it yourself. In Host mode, open the browser window, solve the captcha, and leave the page where it lands. Your assistant picks the task back up from there.
It's worth checking for this whenever a web task seems to stall. A captcha waiting on a human is one of the most common reasons an assistant goes quiet, and it takes you a couple of seconds to clear.