Connecting a browser agent to AI: the assistant really opens the website
Yes. IntraGPT can operate a real browser on our own server. The assistant does not pretend to read a page: it opens it, finds its way around, fills in forms, reads the content and can capture what it sees. The part sits in the platform as beta and we set it up together with the customer.
What the browser agent can do on a website
Really opening, not guessing
The assistant opens the website in a real browser and sees the page the way a visitor sees it. It works with what is there now, not with an old copy or a summary from somewhere else.
Look first, then act
Before clicking anything the assistant takes in how the page is built up: which headings, links, buttons and fields there are. That is why it finds the right spot even when the layout has changed.
Clicking through and filling in
Searching, clicking through, filling in a form, making a choice and submitting: the steps an employee would take too. What it runs into it says in the conversation, so you can follow where it is.
Reading and capturing
The assistant pulls the content off the page and can capture the screen alongside it. Handy when you want to be able to show later what was on that page at that moment.
Signing in where that is allowed
For sites where you want to store sign-in details, the assistant can sign in itself and carry on from there. If nothing is stored for a site it gets no further than the public part and says so plainly.
Work that no longer has to be done by hand
Industry and manufacturing
- Situation
- The lead time for a part is only in the supplier's portal, behind a login.
- What the agent does
- The assistant opens the portal, signs in with the details stored for it, looks up the article number and reads off the delivery week.
- Result
- Work preparation gets the answer in the chat and no longer has to click its way into the portal.
Government
- Situation
- An employee has to consult a public register that is laid out differently for every search.
- What the agent does
- The assistant fills in the search form, submits it and first takes in how the results page looks before reading it.
- Result
- An answer with a captured screen alongside it, so the source stays verifiable.
Legal
- Situation
- A case file needs a record of what a public page said at a particular moment.
- What the agent does
- The assistant opens the page, pulls out the text and captures the screen in the same session.
- Result
- Text and image from the same moment, without anyone cutting and pasting.
Purchasing
- Situation
- A price list sits as a table on a website and changes regularly without anyone passing that on.
- What the agent does
- The assistant opens the page, reads the table and puts it next to what was recorded before.
- Result
- A list of differences in the conversation instead of two screens side by side and a magnifying glass.
How it works in practice
-
1
Switching it on per assistant
The browser function is switched on or off per assistant. If it is off, that assistant cannot open a single website.
-
2
Setting permissions and details
You decide which employees may use the assistant and for which sites sign-in details are stored. Without stored details it stays on the public part.
-
3
What the assistant sees
Only the page it has open at that moment. Every session starts blank and nothing stays signed in afterwards.
-
4
What gets recorded
What the assistant opens and does appears in the conversation and is recorded. Because this part sits in the platform as beta, we look over your shoulder during the first tasks.
Signing in, visibility and where the assistant stops
An assistant that can really click and sign in needs clear boundaries. Those sit in the platform itself, not in a polite request to the model.
Where the data lives: Nederland of de EU
The password stays out of the conversation
Sign-in details you store are used to sign in, but never end up in the chat. The conversation says at most that the assistant signed in on that site.
Only on the site it was meant for
Stored details belong to one site and are used only there. For a site with nothing stored the assistant stays on the public part.
Not for everyone
You decide which assistants have this function and which roles may work with them. For the rest of the organisation the function simply does not exist.
Traceable afterwards
What gets opened and done is recorded. The model runs on our own server in the Netherlands, so what the assistant reads off a page is not sent to an American model provider and is never used for training.
Questions about the browser agent
Does IntraGPT work with a browser agent?
Yes. The assistant can operate a real browser on our own server, and you switch that function on per assistant. The platform labels the part as beta and we set it up together with the customer, so we are watching along during the first tasks.
Does the assistant see my password?
No. Sign-in details you store are used to sign in on the site they belong to, but never appear in the conversation. The chat says at most that it signed in. If nothing is stored for a site, nothing happens and the assistant says it cannot get any further.
What happens when the site changes?
The setup is built for that. The assistant does not follow a fixed script written for one layout, it looks afresh at how the page appears each time. If a button moves or a field gets another name it usually still finds it. If the whole order of steps changes it reports that it cannot work it out rather than clicking something at random.
Can I connect a browser agent to ChatGPT?
ChatGPT can fetch pages, but it is not your browser and not your session, so signing in to a supplier portal is out of the question. With IntraGPT the browser runs in our own setup, you decide which assistants may use it, and sign-in details come from a place you fill and can empty again yourself.
What is the difference with the Chrome extension?
The browser agent runs its own browser on our server and starts every session blank. The Chrome extension uses the employee's own browser, with the sessions they are already signed in to. For the public web and for portals whose sign-in details may be stored, the browser agent is the right choice.
What does beta mean for this part?
The platform labels the browser agent as beta. That means we set it up together with the customer: we pick the first tasks, watch how the assistant moves around those pages and expand once it runs well. The settings are no different from the rest of the platform, the guidance is.
Does the data stay inside?
The browser runs in the same setup as the rest of the platform and the model runs on our own server in the Netherlands. What the assistant reads off a page is not sent to an American model provider and is never used for training. What was retrieved also stays traceable.
What does a browser agent cost?
There is no separate price tag. The work sits in the project we run together: how many sites are involved, whether signing in is needed, which details get stored for that and which employees may use the assistant. We make that concrete in a thirty minute call.
How long does setup take?
Switching the function on takes minutes. The time goes into choosing the first tasks and recording the sign-in details per site. For a portal with a simple login that is an afternoon. For a site with two-step verification we first ask whether the assistant should be going there at all.