Browser Use Agent
The Browser Use Agent is a built-in AI agent that can interact with web pages in your browser tabs. It reads page content, clicks elements, fills in forms, and extracts structured information.
The browser agent always runs on BrowseWiz's own model and uses your credits, whatever model you chose for the chat. If you run out of credits, the chat shows the out-of-credits message and stops.
Turn it on or off
The browser agent is on by default. To turn it on or off, use either:
- the "Enable browser agent" switch in the chat's Tools menu;
- the "Browser automation agent" switch in Options → Tools → Built-in Tools.
Both switches show the same setting. Turning it on in Options first asks for permission to manage your tabs. If you don't allow it, the switch stays off.
Chrome's warnings
When you install BrowseWiz, Chrome lists what it can do, including the two lines below. When an update asks for something Chrome warns about, Chrome lists it and turns BrowseWiz off until you accept it. Accepting turns BrowseWiz back on.
- "Read and change all your data on all websites." BrowseWiz asks for access to all websites at install, so it can open, read and act on whichever site you name without asking again for each site. Its page script also runs in the web pages and frames you open, so BrowseWiz can see the text you select, read the page when asked and show its own notices on the page. A page's content leaves your computer when you send a message with that page or your selection attached, or when the agent, a list read, a scheduled task or a monitor reads it. It then goes to AI providers, and is saved with your chats while chat backup is on. The Privacy Policy lists every place it can go and how long it is kept.
- "Access the page debugger backend." BrowseWiz uses Chrome's debugger for one thing: it tells a background tab that it has focus, so the page keeps drawing while you use another tab. It does this in the tabs BrowseWiz opened itself, while it works in them (see Pages in background tabs).
While BrowseWiz keeps a tab drawing, Chrome shows a bar saying that BrowseWiz started debugging this browser. The bar is at the top of every tab in every window, not only above the tab BrowseWiz works in. BrowseWiz still uses the debugger only in the tabs it opened itself. Scheduled tasks and monitors open their own tabs, so the bar also shows while one of them works on a page, even when you aren't using BrowseWiz. Each monitor checks its page about every 5 minutes, so with several monitors the bar shows often.
If you press Cancel on the bar, the agent carries on, but the tabs it was working in go back to normal background tabs, so it may see an empty or unfinished page. The bar comes back the next time BrowseWiz keeps a tab drawing, such as a tab it opens after that or its next task. Cancel doesn't stop later scheduled runs or monitor checks from showing it. To stop a task or monitor from showing it, pause or delete it in the scheduled tasks drawer. Turning the browser agent off doesn't stop them.
Capabilities
The agent can perform the following actions on a tab:
- Read page content — read the text of a web page, including embedded frames
- Find elements — find links, buttons, inputs and other interactive elements by their words or type
- Click elements — click buttons, links, and other interactive elements
- Input text — fill in form fields and search boxes
- Choose an option — pick an option in a dropdown by its text
- Scroll — scroll the page or a list, for example to load more results
- Press Enter — for example, to submit a search or a form
The agent works from the page's text and elements, not from screenshots. It cannot upload files.
Downloads
The agent can click links and buttons that download files, such as invoices, statements or exports. To know which files were really saved, it needs your permission to see downloads: allow it in Settings → Tools, or with Let BrowseWiz see downloads on a plan card whose steps download files ("BrowseWiz can confirm each saved file only with this permission."). Approve and Reject work the same whether you allow it or not.
With the permission, each click that starts a download waits until the file is saved, up to a minute, so clicking several download buttons in a row saves every file. The agent then lists the files by their real names and sizes. If Chrome blocks a site's repeated downloads, the agent says so; allow them for that site in Chrome's Site settings → Automatic downloads.
Without the permission, the agent reports a file only when the link itself names it, and never claims a download it didn't see.
When the agent reads a page, lists its elements, or gets the result of a click, typing or Enter, it reads fields the page marks as password, one-time-code or payment card fields as hidden, so their values are left out of what is sent to the AI model. BrowseWiz goes by the page's own marking: a field of type password, or an autocomplete value for a password, a one-time code or card details. A field the page doesn't mark this way is read with its value. For example, a password field after a "show password" button turns it into a plain text field, or a code box the page left unmarked. Text the page shows outside such a field is read like the rest of the page, for example a code, password or card number in a status or error message, a summary or an email. Text you ask the agent to type is part of your request: the chat's model, the agent and the safety check all get it, also when it goes into a hidden field.
Tabs
While the browser agent is on, BrowseWiz can also open new tabs in the BrowseWiz Tabs group, load pages in them, list them, and find your active tab. Each agent task works in one tab. During a plan you approved, tasks on different sites can run at the same time, each in its own tab.
Opening and loading pages in these tabs needs no approval; the agent's actions on them do (see Approvals). New tabs open in the background, so the tab you are on stays in front. BrowseWiz does not close the tabs the agent works in during a chat; the background tabs it opens to read a list or set up a monitor close once they are read. A scheduled task run closes the tabs it opened when it ends.
Pages in background tabs
Chrome stops drawing pages in tabs you can't see, and many shops show only an empty page until they are drawn. So that the agent sees what you would see, without taking over your screen, BrowseWiz keeps the tabs it opened itself drawing while it works in them. That covers chats, scheduled tasks, monitors and Claude Code. When it finishes, the tab goes back to being a normal background tab.
This uses Chrome's debugger, which is why Chrome shows a debugging bar while BrowseWiz works in these tabs. See Chrome's warnings for what the bar's Cancel does.
Tab access
The first time BrowseWiz needs to work with your tabs, the chat shows "BrowseWiz needs permission to manage your browser tabs to do that." with a "Grant tab access" link. Click it and allow the permission in your browser's prompt. The chat then shows "Tab access granted. Continuing your request…" and carries on. You only need to do this once. If you refuse the permission, the chat shows "Tab access was not granted. Browser automation stays unavailable." and you can click "Grant tab access" again.
Tabs in the BrowseWiz Tabs group need no access question. Before the agent acts on any other tab, or when BrowseWiz needs the address of the page you're looking at (for example for "this spreadsheet"), the chat asks once: "BrowseWiz wants to work on domain in this tab: page title." It has two buttons:
- Allow for this chat lets BrowseWiz work on this tab for the rest of the chat. The agent adds the tab to the BrowseWiz Tabs group when it first acts on it, and its actions on domain need no approval until you start a new chat, the same as approving with "Also approve later steps on domain this chat" ticked (below). Actions the safety check flags still ask you first. For a page without a domain, such as a local file, the button is Allow.
- Deny refuses access. The agent does not ask again for that tab until you start a new chat.
This question is asked once for each of your tabs, even when approvals are bypassed. Allowing one tab doesn't cover your other tabs, even on the same site. Scheduled task runs only use tabs they opened themselves, never your tabs, even ones you allowed here.
Approvals
Before each action the agent takes on a page, such as reading it, finding elements, clicking, typing, choosing a dropdown option, scrolling or pressing Enter, the chat shows a card that starts with "Waiting for your approval", such as "BrowseWiz wants to read shop.example content." Click the site's name in the card to switch to that tab.
On a click, typing or Enter card, the name of the element, such as element "Search" or "Message" field, is a link too ("Show name on the page"). Click it to switch to the tab and scroll to the element, which is outlined for a few seconds. If the element is no longer on the page, the link only switches to the tab.
The card has these buttons:
- Approve runs this action.
- Approve with Also approve later steps on domain this chat ticked approves this action and every later action on that domain until you start a new chat. The checkbox sits above the buttons, unticked, and is shown only when the action is on one domain and the card has no "Safety check: …" line.
- Deny stops the agent.
The ⋯ button at the end of the card opens more choices:
- Auto-approve all actions on domain approves this action and every later action on that domain until you start a new chat. It is shown only on a card with a "Safety check: …" line whose page has a domain; on other cards, use the checkbox.
- Auto-approve all approves this action and every later action until you start a new chat. BrowseWiz first asks you to confirm ("⚠️ Bypass all approvals this chat session?"). Actions the safety check flags on other sites still ask you first.
- Always auto-approve on domain adds the domain to your Auto-approve domains. Like the other domain choice, it is shown only when the page has a domain.
- Always auto-approve all turns on "Always bypass all approvals" straight away, without asking you to confirm.
The two "Always" choices run the action straight away, then change your settings and show "Setting saved" with an "Open options" link, where you can change them back. If the setting can't be saved, the action still runs and you see "Setting not saved" instead, with the same link to set it there.
When no approval is asked
The agent runs its actions without showing the card when:
- the action is part of a scheduled task run;
- you approved a one-off plan with "Approve & run", also for your follow-up messages in the same chat, until you click Stop, approve or reject another plan, edit and resend a message sent before the plan, or start a new chat;
- "Always bypass all approvals" is on;
- the page is on a domain in your Auto-approve domains;
- you approved with "Also approve later steps on domain this chat" ticked, chose "Auto-approve all actions on domain" or "Auto-approve all", or allowed a tab on domain with "Allow for this chat", earlier in this chat;
- the page has no domain, such as a blank tab or a local file.
The safety check can still ask you in all of these cases.
Approval record
Each approval you give is recorded with your account and kept for 90 days, as the record of what you authorized: approvals you give on a card or on the "BrowseWiz wants to work on …" question, the plans you approve (their title, steps, schedule, what a monitor watches and the changes a scheduled run may make), and turning on "Always bypass all approvals". Actions that run without a card, because of a setting, an approved plan, a scheduled task run or an earlier choice in the same chat, aren't recorded; the plan or setting you approved is. Text typed into fields the page marks as password, one-time-code or payment card fields is recorded as [redacted], and page addresses are recorded without their query string.
Unless you turn off Use my approved actions to improve BrowseWiz in Settings → Privacy, the BrowseWiz team also studies these records, counted across users, to learn which sites and kinds of tasks the browser agent is used for. It uses the site, the kind of each action, the name of the page element it acted on and the title and steps of the plans you approved, never the text you approved typing, and reports a site only when at least 3 accounts used the agent on it.
Settings
Open Options → Tools → Built-in Tools and click "Configure" next to "Browser automation agent". The "Configure Browser Automation" dialog has:
- Always bypass all approvals: skip all confirmation dialogs for browser actions. It is off by default. Turning it on asks you to confirm with "Enable bypass".
- Auto-approve domains: actions on these domains are approved without asking. By default the list holds
google.comand*.wikipedia.org. An entry like*.example.commatchesexample.comand all its subdomains. An entry without*.matches only that domain, sogoogle.comdoes not covermail.google.com. A leadingwww.is ignored. With an empty list, no domain is approved automatically.
Click "Save" to keep your changes.
Be careful when adding domains to the auto-approve list or turning on "Always bypass all approvals". The agent will take actions on these sites without asking for confirmation.
Safety check
The safety check is always on and has no setting. Before every click, text input or Enter key press, and before opening a page on a site you haven't approved (see Opening pages), BrowseWiz checks the action against what you asked for, or the plan you approved:
- it is on the right site, not a lookalike or a site unrelated to your request;
- it has no lasting effect you did not ask for, such as sending, paying, deleting, publishing or changing account settings. Searching, filtering, sorting, paging, closing a cookie banner and changing a cart count as no lasting effect;
- for typed text, it doesn't send personal data, passwords or other credentials, or content from other pages to a site that shouldn't get it.
The check then allows the action or asks you. You decide every action it flags, even one it judges clearly risky.
- Asks you. The approval card is shown, even on an auto-approved domain, with "Always bypass all approvals" on, or during a plan you approved. The card has a "Safety check: …" line with the reason in place of "Waiting for your approval": "this may not be the site you meant", "this has an effect you didn't ask for", or "this may send information the site shouldn't get". Approve runs the action; Deny doesn't.
Approving a card with a "Safety check: …" line:
- Auto-approve all actions on domain also approves later actions on that domain that the check asks about, and Auto-approve all approves them on the sites this card's actions are on, until you start a new chat. On any other site, the check still asks you. The "Always" choices and your saved settings never approve an action the check asks about.
- Any approval on the card also remembers that exact action, so the same action isn't asked about again in this chat. Typing into a password field is never remembered.
If the check cannot run, the chat asks you with the usual card, without a "Safety check: …" line. In a scheduled task run, and when Claude Code browses through BrowseWiz, nobody can answer, so an action the check would ask about, or can't check, is refused.
Opening pages
BrowseWiz opens only http and https addresses. It refuses local files, browser pages such as chrome://settings and other kinds of links. An address without https://, such as otodom.pl, gets it added.
Before it opens a page, the safety check looks at the page's address: that it's the right site, and that the address doesn't carry personal data or other information the site shouldn't get. A site isn't checked when it is:
- the site of a step in the plan you approved;
- named in one of your own messages, as an address or a domain such as
otodom.pl; - in your Auto-approve domains;
- one you approved earlier in this chat with "Also approve later steps on domain this chat" ticked or "Auto-approve all actions on domain", or with "Auto-approve all" on a card with a "Safety check: …" line for an action on that site.
Otherwise, "Auto-approve all" and "Always bypass all approvals" don't approve sites to open.
- Asks you. The card says "BrowseWiz wants to Open address." with its "Safety check: …" line. Approve opens the page, and the same address isn't asked about again in this chat. Deny doesn't open it. Auto-approve all actions on domain also opens later pages on that site without a check until you start a new chat.
- If the check can't run, the same card asks you, without a "Safety check: …" line, and approving it doesn't stop the address being checked next time.
In a scheduled task run, a page on a site that isn't in the task's plan or your Auto-approve domains is opened only when the check allows it.
Multi-step tasks
For a bigger request, such as one with many steps, more than one site or work that repeats, BrowseWiz first shows a plan card with the steps and waits for you:
- Approve & run starts the task. The agent then runs the plan's actions without approval cards, and keeps doing so for your follow-up messages in the same chat ("now the second one"), until you click Stop, approve or reject another plan, edit and resend a message sent before the plan, or start a new chat. The safety check still applies.
- Reject cancels the plan.
A plan with a schedule has Save and run now and Save in place of Approve & run. Both save it as a scheduled task: Save and run now also starts one run of it as soon as BrowseWiz's reply ends, and Save leaves the first run to the schedule. A plan that changes an existing scheduled task shows "Updates scheduled task" with the task's name, and has the same two buttons.
While a plan card waits it reads "Waiting for your approval", and "To change the plan, type what to change." under its buttons. If you send a message while it waits, the plan is rejected and your message is sent, so you can say what to change. If you click Stop, the card is marked "Cancelled".
Reading a whole list into a spreadsheet
When you ask for every item of a list in a spreadsheet, such as all the results of a search, and the list's pages are linked by a "next" link, BrowseWiz reads the list with code instead of having the agent read each page. It needs the browser agent and the Spreadsheet tools switched on. A one-time read needs no plan card.
- BrowseWiz looks at the list page and works out where each item's values are: its name, price, link and the other columns you asked for.
- It reads the first page and checks the rows. When a value is missing or wrong, it corrects where the value is read from and reads the page again.
- It then reads the pages one after another, following the list's own "next" link, and writes the rows to the spreadsheet.
The AI model isn't called for each page. It gets an outline of the list page, the first three rows read, the first item's ID attributes, how many rows have each value and, for up to five columns where a value is missing, the text of the first item that lacks it. When a read or a write fails, it also gets the item or the values that caused it. The other rows go from the pages to your spreadsheet without passing through the AI model. While it reads, the chat shows "Using read a list..."; a long list can take a few minutes, with no progress shown. A "Read a list" row appears when the read ends.
- Tabs. Each page opens in a new background tab in the BrowseWiz Tabs group and is closed once it's read. Pages load one at a time, with a pause of 3 to 8 seconds between them. Before the read, its first page's address is checked as when BrowseWiz opens a page (see Opening pages); the first look at the page in step 1 loads it without that check. The pages after it must be on the same site (with or without
www.). BrowseWiz only loads and reads these pages, so there are no approval cards for actions. - Limits. One read covers up to 20 pages and 1,000 rows.
- Values are copied as the page shows them, text up to 300 characters and links up to 2,000, and each row's Source URL is the page it was read on. In a number or currency column the number is written: "3 999,00 zł" becomes 3999 in a PLN column. A value the page doesn't show for an item is left empty.
- An item listed twice. An item that reads the same in every column is kept once. Some sites change the order of a list between its pages, so they show some items twice and leave as many others out. When the page states a total and the pages showed fewer items, the AI is asked to tell you how many items the pages didn't show.
- A page that can't be read. When the site blocks a later page, or it takes too long to load, isn't found, shows no items or is on another site, the read stops there, keeps the rows of the pages before it and is reported as incomplete. When the first page can't be read, nothing is written. When an item on any page lacks a value every item must have, the read fails and nothing is kept; BrowseWiz then corrects where the value is read from and reads again.
- What it can't read. A list that loads more items as you scroll, or whose "Load more" or next control is a button and not a link, and values that need converting or judgment, such as dates. BrowseWiz uses the browser agent for those.
- Stop. The stop button ends the read, closes its tab and keeps nothing of that read. While rows are being written, it ends the write after the group of 50 in progress; the rows already written stay in the spreadsheet.
- Scheduled tasks. A plan step that reads a list keeps the list's address and the places each value is read from, so every run reads the same columns. If the site changes and the read fails, the run is told to use the browser agent instead.
The caution under Example Use Cases of the Spreadsheet toolset applies: check that the site's terms allow it, and don't copy all or a large part of a site's listings.
While it works
While the agent works, the chat shows "Using browser automation agent...". Each agent task then appears in the chat as a "Browser automation agent" row with the task in quotes. Click the row to see the task BrowseWiz gave the agent and its result, with a copy button.
Each task has a limited number of steps. If the agent runs out of steps, it reports what it finished.
To stop the agent, click the square stop button in the message box. This also removes any open approval card and ends an approved plan run.
Example Use Cases
- Web research: "Search for 'best project management tools' on Google and return the top 5 results"
- Form filling: "Fill in the contact form on this page with the following details…"
- Data extraction: "List all article titles on this Wikipedia page"
- Workflow automation: "Search for flights from NYC to London on this travel site"
Tips
- Keep requests focused and specific — describe exactly what data you need or what action should be taken
- BrowseWiz gives the agent one clear goal per task (e.g., one search query, one form to fill, one page to read) and splits bigger requests into several tasks
- Each agent task gets only the instructions BrowseWiz writes for it, not the whole chat — name the exact values, fields and pages in your message when they matter
- The agent works in your browser, with the sites you are signed in to — if a site asks you to sign in, the agent stops and tells you; sign in yourself, then ask again