How to fill a carrier portal with the Claude Cowork browser, step by step
- Open Claude Cowork on your computer. This runs in the Claude desktop app. You cannot do it from claude.ai in a web browser, so open the app on your computer first.
- Find the browser button at the top. There is a small browser button at the top of Cowork. That is the whole feature: Claude can open any website and take actions on it for you.
- Give Claude the information to work from. Drop in what you already have. In the demo that is a policy declarations page PDF. A screenshot from your AMS, typed notes, an old policy document, or a mix of all of them works the same way. Treat it like an assistant and give it as much as you can.
- Paste the portal link and say what you want. The request in the video is simple: use your internal browser to open this website, if you need me to log in let me know and I will put in my login, then fill out the form to the best of your ability without making anything up based on the fields it asks for.
- Sign in yourself when Claude asks. Claude opens the portal in the pane beside the chat and stops at the login, telling you to go ahead and sign in. You type the credentials into the portal yourself. Check remember this device so you are not doing it every time.
- Tell Claude you are in. Say you are logged in, the same way you would tell a coworker. Claude confirms it can see the session and starts filling.
- Watch it fill, and read what it flags. Fields populate in real time: named insured, business name and DBA, email, requested effective date, renewal versus new, the roofing, plumbing and heating years, and the coverage amounts. When it finishes it lists what was missing and what it wants you to confirm.
- Check its work against the source document. In the demo the roofing 2021, plumbing 2009 and heating 2016 entries all trace back to the declarations page. Claude also states plainly that nothing was invented, which is a claim you should spot-check rather than take on faith.
- Correct anything wrong, then submit. Ask why it entered something and it explains its reasoning, then fixes it. You can also click into the page and edit it yourself. Claude does not submit until you tell it to.
The whole internet, inside Claude
The new piece is a browser button at the top of Claude Cowork. Click it and Claude has a real browser, which means it can open any website, read it, and take actions on it. Shopping on Amazon, pulling something out of a web app, filling a form: it is a browser, so the range is whatever a browser can do.
For an insurance agency the obvious use is the one that eats the most hours. Carrier portals and quoting sites are where Saturdays go. The information already exists on a declarations page or in your AMS, and someone still has to retype it into a portal, one field at a time.
One requirement is worth saying twice: this is the Claude desktop app only. You cannot do this from claude.ai in a web browser. If you are reading this and want to try it, open the app on your computer.
| Built-in Cowork browser | Claude for Chrome | |
|---|---|---|
| Where the browser is | Inside the Claude desktop app, in a pane next to the chat | Your own Chrome, driven through the Claude for Chrome extension |
| What you install | Nothing extra, it is the browser button in Cowork | The Claude for Chrome extension |
| Signing in to a portal | Claude stops at the login and you type your credentials into the pane yourself | You are already signed in in your own Chrome |
| Watching it work | Side by side with the chat, in real time | In your own Chrome tab, in real time |
| More than one at a time | Yes, as many Cowork tasks as your computer will handle | Yes, across Chrome tabs |
| Where it runs | Desktop app only, never claude.ai | Chrome only |
Give it something to work from
Claude knows the portal once it opens it. It does not know your client. That gap is filled by whatever you hand over at the start, and the format is deliberately flexible.
In the demo it is a policy declarations page dropped straight into the chat. It could just as easily be a screenshot from your AMS, an old policy document, notes you typed out, or a mix of all of them. You can also tell Claude to go get it from your AMS or CRM. As Ben puts it, treat it like an assistant and give it as much as you can so it has everything it needs to fill the form.
That framing is the actual skill. Everything that goes wrong later in the video traces back to a document that was ambiguous rather than to Claude being incapable, which is exactly how it goes with a new hire.
Getting into the portal without handing over your password
Passwords never go in the chat. The instruction given up front is to open the site and say so if a login is needed, and that is what happens: Claude opens the portal in the pane on the right and reports that the page is showing a login dialog, then asks you to sign in.
You type the credentials into the portal yourself, in that pane, the same as you would in any browser. Check remember this device so the next submission skips the step. Then you tell Claude you are logged in and it picks up where it left off.
Before it starts, Claude states what it will and will not do: it will fill the form, screenshot it, and leave the submit to you. That default is the right one for a carrier submission, and it is worth confirming it holds every time rather than assuming.
Watching it fill, and reading what it flags
The fill happens in view. Named insured, business name and DBA, email address, the requested date, renewal versus new business, then the building details and the coverage amounts. In the demo the roofing, plumbing and heating years going into the portal match the declarations page exactly: plumbing 2009, heating 2016, roofing 2021.
The more useful output is the list at the end. Claude reports that the form is filled and verified, that every field traces back to the declarations page, and that nothing was invented. Then it names the judgment calls. Coverage C had nowhere to go because the portal only has A and B, so the business income with extra expense detail went into remarks in full.
It also singles out the one entry it is unsure about. The NAIC code it entered, 99017, is the expiring carrier's code, because that was the only NAIC code on the document, and the submission is going to a different carrier. Claude says outright that the field may want the receiving carrier's own code and that this is the one entry it would have you confirm. A tool that tells you which single field to double-check is more useful than one that reports finished.
When it gets something wrong, ask why
It put the wrong date on the form. Asked why, it explained the mapping: the date came off the declarations page, where section 8 has a signature date, and the labels sat next to each other so it mapped one to the other. Then it corrected itself without being told the answer. The form's date is attesting to the submission, so it should be the day you actually submit, not the day the old policy was signed.
Its replacement date was still off by one, and the reason is worth knowing. Claude was working in UTC while Ben is on Pacific time, so its idea of today was tomorrow. Telling it the time zone fixed it. If a date looks a day ahead, that is the first thing to check.
The page is also still just a web page. Ben clicks into the date field, types the wrong thing by accident, clicks out, and tells Claude to fix the date and submit. You can interact with the site exactly as you would in Chrome, Edge, Safari or Firefox, because that is what the pane is.
The pattern underneath all of this is the one worth copying. Ask why it did something, the way you would ask a human assistant who put the wrong date on a form. You usually get either a correction or a reason, and both are useful.