Skip to main content
A Browser use step has an AI agent open a website for each contact and work toward a goal you write. The contact then moves on down the branch for the outcome it reached.
The Browser use step panel. The goal asks the agent to open their website and start a free trial with the sender's email. Outcomes Trial started, Signup blocked, and Needs human sit below it, then a Password value with Hide value on, a line naming the sender it acts as, and the switch that lets the agent read the sender's inbox turned on. A price line gives the most a contact's run costs, the price per browser minute, and the longest a run lasts. Under it, the Trial section holds a Try on one contact button.

A Browser use step's panel, filled in from a starter goal.

​
Add a Browser use step

1

Open the step picker

Open a sequence in Sequences and go to Builder. Point at the gap between two steps and click Add a step here.
2

Choose Browser use

Under Browser, click Browser use. Its panel opens beside the canvas.
3

Write the goal

Under Goal, say what the agent should do on the site. Type / to add a column, such as the contact’s website. Never put a password in it.
4

Name the outcomes

Under Outcomes, rename Done, and click Add outcome for each other way it can end. Each outcome becomes a branch on the canvas.
5

Add steps to each branch

Add what comes next under each outcome, and under Needs human.
With no goal yet, the panel offers Fill in their contact form, Start a free trial on their site, and Sign up to this directory. Each starter fills in the outcomes and values too.

​
Know what Needs human does

Every Browser use step has Needs human, the last outcome, which you cannot rename or remove. It never waits for a person: the contact goes straight down that branch and gets a task in Tasks.

​
Save values to the table

Under Saved to the table, click Add value for each thing the agent should bring back, such as a username. Each value gets its own column on the audience table. Turn on Hide value for a password or anything private. Its cells read ”•••••• (open the contact’s history to reveal)”.

​
Know who it acts as

The agent acts as the sender of the mailbox that emails each contact, so the sequence needs a sender with a connected mailbox.

​
Let it read sign-up emails

Turn on Let the agent read the sender’s inbox when a site confirms an account by email. The agent then reads the code or link the site sends. While the switch is off, a starter that needs it marks it Needed for this goal.

​
Try it on one contact

1

Open the step

Click the Browser use step on the canvas, then click Try on one contact in its panel.
2

Pick a contact

Choose someone from the audience. The dialog shows who the agent acts as and the most the run can cost.
3

Run it

Click the button that names the most it can cost, such as Try it · Up to 56 credits. The panel shows the trial as it runs.

​
Watch what it did

The Audience tab’s Outcome column and filter show each contact’s latest outcome. Click a contact to open their history: its outcome and why, the saved values, Steps, and Watch replay.

​
Good to know

Yes. On the Sequences page, click Browser use under Ways to start, or under New. It asks for the audience first, as every way to start does. The draft holds one Browser use step with no goal yet, which ends on Done or on Needs human.
Clear the goal and pick another starter, and it replaces the outcomes and values the first one left as they were. An outcome you renamed, or one with steps on its branch, stays. So does a value you renamed or added yourself.
A starter opens each contact’s website. When no audience column has “website” or “domain” in its name, the panel asks Which column holds their website? and points the goal at the column you pick.
A trial runs the saved step once. It never moves the contact, writes a cell, or counts in the step’s results. The panel shows the latest trial with its values, Steps, and Watch replay. It is charged like one contact’s run: by the browser minute, and nothing if its browser never opens.
No. Each step runs one trial at a time. While one runs, Try on one contact is greyed out, and pointing at it reads “The last trial is still running”. A trial waits for a free browser for up to an hour, then ends without a charge.
Anyone a run of the sequence could reach. A trial refuses a contact outside the audience’s view, one stopped on this sequence (an unsubscribe stops them too), one on Do not contact, and one inside your workspace’s cooldown. A trial in line ends without a charge when its contact is stopped or put on Do not contact before it starts.
A sequence whose steps are all Browser use sends no email, so it has no reply rate. When every sequence listed is one, Reply rate becomes Runs: the runs that ended, and how many of them went down Needs human. Positive becomes Credits, the credits its runs were charged. In a list that mixes them, a Browser use sequence reads ”-” in both.
Its row counts its runs, how many went down Needs human, and the credits they used. It sends nothing, so it never counts as a send. While every contact moving is on a Browser use step, the headline reads Browser runs in progress.
The step’s card on the Builder tab and the Audience tab’s header count them, such as “12 done · 3 browsing · 10 waiting”. A contact waiting for a browser, or held while Browser use is unavailable, counts as waiting.
It is priced per browser minute, up to a limit per contact. The step’s panel shows the most a contact’s run costs, the price per browser minute, and the longest a run lasts. The editor header counts the step in the sequence’s estimate.Each run is charged for the minutes it used, rounded up to a whole minute. The browser opens only when the agent first needs a page, so a run that ends before then costs nothing, such as one whose contact has no web address to open. See Credits and billing.Each run in a contact’s history shows what it was charged and the browser minutes it used, and a trial shows the same.
No. It takes free actions only, never enters card details, and never buys anything. When a site asks for payment, the contact goes down Needs human.
Open it in Tasks to read the reason, then open the contact’s history or watch the run’s replay. Do the rest by hand, then click Mark done in its pane.Marking it done changes nothing for the contact, who has already moved on. A trial never leaves a task. When a contact sent through the step again goes down Needs human a second time, the task opens again with the new run’s reason and replay.
When the agent chooses it, or when the run ends before the agent could choose:
  • A CAPTCHA it cannot pass, a code sent by phone or text message, or an identity check
  • A page asking for payment
  • A page asking for a personal detail it was not given, such as a phone number
  • No page confirmed the goal was reached
  • The run used all its minutes, or reached its spending limit
  • The workspace ran out of credits, which also pauses the sequence until there are more
  • The run stopped because it was no longer needed, such as when the contact left the sequence
  • The browser stopped responding, or the AI model failed
  • The run stopped before it could open a browser three times
The reason reads under the run in the contact’s history. Its chip is amber when the agent chose Needs human, and red when the run broke off. If nothing is connected to Needs human, the sequence ends for that contact.
The goal reads a column that is empty on their row. The run never opens a browser and costs nothing. The contact reads Failed, their history shows the step with what is missing, and the banner on the Audience tab names it, such as “Ben Blank has no Website.”Fill the value in on the audience table, then select Retry. While it is still empty, Retry leaves the contact failed and says which value is missing.
A contact nothing has emailed yet gets one at the step, for good, and their emails come from it too. It is the connected mailbox of the sequence’s senders that already emails the fewest contacts. A contact already written as one sender, or worked by their LinkedIn account, gets that sender’s mailbox.A contact whose mailbox was deleted, or belongs to a sender who is not on the sequence, stops there, as at an email step. One whose mailbox is on no sender waits until it is back on one. One whose mailbox is disconnected still runs, unless the step reads the inbox.A contact on Do not contact, or inside your workspace’s cooldown, stops at the step before a browser opens, and again if that happens while their run waits in line. A hard bounce alone does not stop a run.
A run’s history lists only the values it looked for. Nothing found means it looked and found nothing. A value it never got to is left out.
Click the line under Steps. A line too long for the panel ends in an arrow, and a click opens it in full. The reasons it gives read as what the run did and what the pages showed.
No. Every run starts in a fresh browser, with no sign-ins and no memory of earlier contacts or earlier sites.
The sender’s name, email address, job title, and company. Never the contact’s own email address or phone number, and never a detail it was not given. Any account it creates gets a new password, so save a Password value with Hide value on to keep it.
No. Never put a password in the goal: the goal is sent to the AI model and kept in the sequence’s Change history, where anyone in the workspace can read it. The step has no field for one either. A site that asks the agent to sign in with details it was not given sends the contact down Needs human.
Only mail from the site open in the browser that arrived after the run started. It never reads mail from sign-in or email services, and it uses a code or link only on the site that sent it.It reads the mailbox that emails this contact. While that mailbox is disconnected, the contact waits for it.A site that had already emailed that mailbox before the run began is never read at all. The sender may hold an account there, and a sign-in code reads the same as a sign-up’s. The agent cannot finish the sign-up, so the contact goes down Needs human.
Open the contact’s history on the Audience tab and click Show hidden values. It reveals every hidden value of that run at once, and Hide values covers them again.Each reveal is recorded in the sequence’s Change history, in the actions menu, with the values’ names and the contact’s, such as “Showed Password for Satya Nadella”. The values themselves are never recorded. Opening the replay of a run that kept a hidden value is recorded the same way, because the recording shows what the agent typed.A hidden value can’t go into an email, a prompt, or a goal. The ”/” menu leaves it out, and a step that reads it can’t be saved.
On the sequence’s Activity tab. Each finished run is a row with the contact, the outcome it ended on (or Needs human, or Failed), the site it worked on, and when. Click a row for why it needed a person or failed. A trial of the step is never listed there.
In a column named after the sequence, such as “Pipeline outreach: Outcome”, one for each Browser use step. A run writes the outcome it ended on, or “Needs human”.Renaming or removing an outcome keeps the column, and every cell keeps the name the outcome had when that run ended. Formulas and AI columns can read it like any other column.
Contacts who already took it keep it in their history. For contacts still on the step, the agent chooses among the outcomes that are left. A run already under way ends on Needs human, because the outcomes changed while it ran. Review changes says how many of each there are before you make the change live.
The column is named after the sequence and the value, such as “Pipeline outreach: Username”. Renaming the value renames its column. Removing the value keeps the column, and everything earlier runs saved in it.A hidden value’s cell holds only ”•••••• (open the contact’s history to reveal)”. The value itself stays in the contact’s history. A run that found no value leaves its cell empty.
Because another column reads it. A formula, an AI column, or a run condition that read the value keeps a copy of it, in the open.Take the value out of that column first, then turn on Hide value. The save names the column that reads it.
Up to 8 outcomes besides Needs human, and up to 8 values. Removing an outcome deletes its branch, and when steps are on it, a confirm asks first and deletes them too. On a sequence that never launched, a removed outcome’s slot is free again for any new outcome.
There is no limit, for your workspace or overall. Every contact that reaches the step is started, and one waiting for a free browser reads Waiting for browser until it starts on its own. A wait never counts as a failure and costs nothing. A contact whose run is going reads Browsing.
The contact waits, reading Browser unavailable, and the run starts on its own once Browser use is back. It tries again after 2 minutes, then waits twice as long each time, up to 30 minutes between tries. Nothing is charged while it waits, and the Audience tab says how many contacts are waiting.
A run whose browser is open stops within a few seconds, and the minutes it ran are charged. The message after Pause sequence says how many browsers stop. A run that stopped before it changed anything on the site runs again after you resume. A run that already reached an outcome keeps it, and the contact goes on by that outcome once the sequence resumes.
Not once its browser opened, because the site may already hold an account from that run. Send the contact through the step again yourself if you need a second try. A run that stopped before its browser opened reached no site, so it goes back in line on its own, waiting a little longer each time. After its third such stop, the contact goes down Needs human.
For a limited time. After that, Watch replay opens on This replay has expired. A run whose browser never opened shows Nothing was recorded, and a run across several pages has a tab for each page.The replay’s title names the contact, the site, and the day, such as “Replay · Satya Nadella · acme.com · Nov 3, 2026”.
The launch checks list each Browser use step with an empty goal, with Write the goal beside it, and its card on the Builder tab reads “No goal”. A starter goal that confirms the sign-up by email gets Turn on inbox access while the switch is off. When the audience table has no room for the step’s outcome and value columns, Remove columns opens the audience table first. See Clear the launch checks.
It says how many websites the agent visits and as which senders, and the columns the launch adds for the step’s outcome and values. When some contacts lack a column the goal requires, such as a website, it says how many will fail at the step. That line is advice, so Launch stays enabled. With no sender yet, its Senders row says “Browser use acts as each contact’s sender” beside Add a sender.

​
Next steps

Build a sequence

Add steps and branches to the canvas.

Branch on a condition

Split a branch on what you know about the contact.

Clear the launch checks

Fix what blocks a sequence from going live.

Credits and billing

See your balance and buy credits.