# Agent on your websites (/docs/en/guides/agent-websites)

The Agent can work on the websites your practice already uses online, such as insurer portals and referral networks. It works in your own browser, in a new tab or in one you already have open, fills in what the patient record supports and asks you what it cannot know. Nothing is submitted to a website until you approve it.

It works through the **Salisan extension** for Chrome and Edge on a computer. The extension only connects your browser to Salisan: you keep talking to the Agent in Salisan, in the **Agent** page or its side panel. Phones cannot run the extension. If you use the Salisan desktop app, the websites open in Chrome or Edge while you keep talking to the Agent in the app: see [From the desktop app](#desktop).

## Choose where the Agent works [#websites]

Open **Settings**, choose **Agent**, then **Websites**. The practice makes one choice:

- **All websites**: the Agent can work on any website. It asks before it sends anything.
- **Only these websites**: the Agent can work on the websites you list and no others. Type an address, for example `vecozo.nl`, and choose **Add**. Its subdomains are included.
- **Off**: the Agent does not work on websites.

Changing this needs the **Manage AI credits** permission, which Owner and Admin have; everyone can read it. There is nothing to set per website: whatever the Agent may open, it may read and fill in, and submitting always waits for your approval.

Under **All websites** you see the websites the Agent has worked on, with the short notes it keeps about finding its way there. Notes never hold patient details. An address on your own network, such as software that runs in the practice, is open to the Agent only when it is on your list: add it under **Only these websites** first.

## Add the extension [#extension]

1. Install the **Salisan** extension from the Chrome Web Store, or from Microsoft Edge Add-ons for Edge.
2. Open Salisan in the same browser. **Settings → Agent → Websites** shows **Extension connected in this browser** when it is ready.

There is nothing to allow per website: the Agent opens and reads what it needs, and asks you only before it sends something.

Sign in to each website yourself, as you always do. The Agent never sees passwords, codes, smart cards or authenticator apps, and Salisan does not store them.

## From the desktop app [#desktop]

The desktop app cannot run a browser extension. You keep talking to the Agent in the app, and the websites open in Chrome or Edge on the same computer.

1. Add the extension to Chrome or Edge, as above.
2. Open Salisan in that browser and sign in. The first time you ask the Agent for website work in the app, the chat offers **Open in my browser**, which opens it for you.
3. Leave that Salisan tab open while the Agent works. Its tabs open next to it, in the **Salisan** tab group.

You approve in the app or in the website's tab, whichever you are looking at. When a website needs you to sign in while you are watching the conversation in the app, its tab comes to the front in the browser.

## Ask the Agent [#ask]

Ask in your own words, for example "Request patient transport for Sanne de Vries on VECOZO", or "finish the form in the tab I have open". The Agent opens a tab of its own, or works in the tab you pointed to, where you left it. It never goes through your other tabs.

You can always see where it is:

- Every tab the Agent works in sits in a **Salisan** tab group in your browser. A tab of yours joins the group while the Agent is in it.
- In the chat, a line above the message box names the tabs it is working in. Choose one to bring that tab to the front.
- In the tab, a thin frame marks the page, the Agent's pointer moves to what it clicks and types, and the fields it filled keep a thin outline.
- Each step appears in the chat as it happens, named the way the website names it: the button it clicked, every field it filled and what it put there. Identifiers from the record are shown masked.
- When the Agent changed something on a website (a form it filled in and left for you, a saved draft, a send the website confirmed), its answer ends with a picture of the page. Values from the record are covered in it, and the Agent itself does not see it. The picture is kept with the conversation and deleted with it.

While it works:

- When a website asks you to sign in, the Agent waits until you are in. If you are watching the chat, the tab comes to the front; if you have moved on to other work, Salisan tells you instead and shows the tab when you ask. If the website signs you in on a separate service, you are taken there. The Agent does not read those pages.
- When the record does not answer a question on the form, the Agent asks you in the chat.
- When a form needs a file, the Agent asks you to attach it in the tab.
- A national identifier (BSN, KVNR, AHV), an insurance number or an exact date of birth goes from the record straight into the form. It is never sent to the AI, and it cannot leave for any other website than the one the Agent is filling in.
- You can leave the chat and work elsewhere in Salisan; the Agent keeps going, and the chat is marked when it needs you or is done.
- When the Agent has waited about 20 seconds for your approval or for you to sign in, and you are not looking at the conversation, Salisan sends you a notification: in your inbox, as a push notification where you have those turned on, and from the desktop app. It also tells you when a longer piece of work has finished. These notifications never name a patient.

To step in:

- Choose **Take over** on the website, press **Esc**, or drag the tab out of the Salisan group. The Agent stops working in that tab, and the chat shows **Hand back** for when you want it to carry on.
- Choose **Stop** in the chat to stop the Agent altogether. Nothing is left waiting on the website. Say "continue" and it picks the form up where it was.
- If you close a tab the Agent was using, it says so and asks before opening it again.

When the work is done, the Agent closes the tabs it opened. A tab of yours is only handed back.

While the Agent works, Chrome shows a bar saying that the Salisan extension is debugging the browser. That is how the extension reads and fills the page. The bar goes away when the Agent is done.

## Approve what is sent [#approve]

Whatever a website's button says, the extension holds anything the Agent's action would submit and shows it to you in the chat. The card stays short: what the Agent is about to send in one sentence, how many fields the form has, and how many of them came from the record or from the Agent.

Looking something up is reading, not sending: a search box or a calculator on a website goes through without a card, unless what the Agent typed includes an identifier from the record or the search goes to another website.

- A short form is shown in full on the card.
- For a longer form, choose **Review all**. A panel opens with the whole form in the website's own sections, and you can send from there.
- **Show page** brings the tab to the front, where the fields the Agent filled are outlined. You can correct a field there; the card then shows the new value as yours.

After each value the Agent put there you see where it came from:

- **Record**: taken straight from the patient record.
- **Agent**: typed or chosen by the Agent.

Values you entered yourself, or that the website filled in, have no word.

Choose **Send** (or **Send and continue** for a step of a longer form), or **Don't send**. If the page changes after you looked, you are asked again.

You can also answer in the website's tab. While a send waits, the bar at the bottom of the page asks **Send this form?** with **Send** and **Don't send**, and they do exactly what the card's buttons do. If the Agent ticked a declaration itself, **Send** in the tab takes you to the card first, so you see that line before you confirm.

Declarations, consents and signatures stay yours: the Agent asks you to tick them in the tab. If it ticked a box itself, the card says so in a separate line, so you can check that it is yours to confirm.

After a send, the chat keeps a receipt. **What was sent** shows the same fields again, also after the tab is closed.

Every send is also kept outside the conversation:

- When the form is about one patient, the card names the patient before you approve. Afterwards the send appears in the activity on their patient record as **Sent to** the website, with the website's reference number when it gave one. Open it to see what was sent and what the website answered.
- **Settings → Agent → Websites** lists the practice's sends under **Sent**, for people who can manage the Agent. Details of a send about a patient show only to people who can open that patient's record.
- Each send is written to the practice's audit log, with who approved it, when, and for which patient.

<Callout type="info">
Some websites save a form while it is being filled in, or look something up as you type. That reaches the website when the Agent types, as it would when you type. What waits for your approval is every submit: a button, the Enter key, a choice that sends the form.
</Callout>

## What the Agent reads [#privacy]

The Agent reads the website as text. It does not take screenshots, and it sends website content to the AI only to work on your request. Text on a website is treated as information, never as instructions to the Agent.

When you ask about your open tabs, the Agent sees the website and title of each tab. Salisan's own tabs are never among them.

Patients you excluded from AI assistance are masked in everything the Agent reads on a website, by name and by their numbers, also when they appear in a list next to the patient you asked about.
