Everybody is sleeping on the claw
extension for Chrome and we shouldn't be
because it basically takes an LLM and it
puts it inside the most popular browser
on the planet where we all spend a lot
of our days and it says, "Hey, do you
want a free agent that can do useful
work on the internet?" And it turns out
if you can define something you do
repeatably, if you can define a piece of
work you don't want to touch, you can
give it to Claude on the internet and
make Claude go do it. And that saves
people, yes, real people, dozens of
hours a week because it turns out that a
lot of us do a lot of stuff on the
internet that just is suffering and
toil, right? Like we just sit there and
we're like, "Oh man, I have to go and
check these three different websites and
pull all the numbers for the weekly
report." Or, "Man, I don't want to talk
to T-Mobile today and talk with their
customer service agent and make them
give me a discount." or man, I really
don't want to sit there and figure out
how I move my 15 meetings this week and
get them blocked into something that is
more manageable. Well, that's what
Claude in the browser is for. So, in
this video, I'm going to dive into what
it is. I'm going to give you specific
examples that build on each other so you
can see how you can get to fairly
sophisticated behaviors. And I'm going
to talk about something that I don't see
talked about enough, which is the
difference between using Claude
extension in the browser, using Claude
in Claude code in the terminal to
navigate in Chrome, which you can also
do, or using co-work, which is a app
that Claude has launched to do agent
actions on your computer and using that
to navigate Chrome. So, there's like
lots of different ways to navigate
Chrome. It can be confusing. We're going
to lay out the differences. We're going
to focus mostly on the browser extension
because most of us are in Chrome. Chrome
is not something you have to install
separately. It just kind of comes with
most computers. We're going to talk
about why it matters. We're going to get
super practical. Let's start with one of
our favorites, letting AI fight your
customer service battles for you. Carl
Votti is a product manager and cloud
code instructor, and this year he posted
a thread that talked about a billing
dispute with AT&T. The reason this
little post about AT&T went viral is
that Carl did not go and sit there for
45 minutes on hold on the phone. He did
not go into the chatbot himself.
Instead, he launched Claude code with a
Chrome extension, opened up AT&T's live
chat, and told Claude what he wanted, a
refund from a recent outage. Claude ran
the conversation. It read the agents
responses. It typed in contextual
replies, and it pushed back when the
initial offer was low. It escalated
politely when the agent stalled. And
eventually, Claude negotiated its way to
a $100 credit. Now, I know what you're
thinking. You're thinking, "This is
Claude code. It's not the same thing.
You have to type into a terminal to get
Claude to use a browser. No, the point
is Claude in the browser works. You can
use Claude from the terminal if you
want, if you're comfortable with that.
You can also use it directly from the
nice extension sidebar and get exactly
the same results. That's why there's a
nice extension sidebar. The underlying
mechanics are exactly the same. Claude
has to read the text on the active page,
type into an input field, click a
button, maybe it has to navigate. All of
that back and forth happens
automatically when you give Claude a
task. Now, to be honest, Carl and I have
both noticed that this is not fast.
Claude is not super quick at getting
this stuff done. And so, this is
something where it will take longer than
it would take a human to do. But the
benefit is really clear. You, the human,
don't have to do it. And so if it's
something where you want some upside,
you want some value, maybe you want a
credit back on your phone bill and it's
worth sitting around and letting the
agent work on it. It's a really simple
use case, it could put money back in
your pocket and you don't have to do
anything. And as I said, like Carl used
code, but the same principle works right
in the browser if you open up a sidebar.
Ultimately, the larger pattern here is
that if you have a chat window in your
browser, that chat window is something
that you can operate with clot. It's not
just AT&T, it's Verizon, it's any kind
of utility company. It's a chatbot with
Amazon. Anything where you can type into
the screen, Claude can now do a typing
for you. And it's as simple as a browser
extension. Now, one of the things I
emphasize is that agents need to be
proactive to work well. And Claude
builds that in in their Chrome extension
from the get- go. This is a capability
that ultimately turns a single clever
trick into useful repetitive work. And
if you're wondering, how does Claude
know what to do repetitively or what
kind of repetitive work would I do? Let
me give you an example to make it come
alive. Let's say that your job is to
pull analytics. When you want to make
Claude do this, you don't type a bunch
of text to say, "Claude, go do this."
Instead, you click the record icon in
the extension panel. You perform the
task you want Claude to learn like
pulling analytics from a dashboard,
checking a competitor's pricing page,
extracting data from a CRM, whatever it
is. Anything that is in the browser that
you can do, Claude will see. Then you
stop the recording. You save that entire
workflow as a shortcut in the Claude
extension. Now you can schedule that
shortcut. You can schedule it to run
daily. You can schedule it to run
weekly. You can schedule it to run
monthly or annually. You just click the
clock icon. You set the cadence and
Claude runs the task on autopilot. It
doesn't need you. It doesn't need a
reminder. It just executes as long as
the computer's up and the browser's
active. And yes, this really works. You
can set up a recurring shortcut to pull
up LinkedIn invitations and look at
who's trying to connect with you. A
recurring shortcut to check your
favorite YouTube videos, maybe like this
one, and get them all listed in links
that you can grab whenever you want
them. You can set up a recurring
shortcut to look at your emails and
extract only the ones you care about.
You can set up a recurring shortcut to
look for all of the new restaurants in
your local neighborhood and make a list
of them that you can go to on a Friday.
Some of this stuff is actually personal
and some of it is professional and it
doesn't matter. The point is if you can
do it on the web and you have to do it
more than once, you can make a claw
shortcut to do it. And how many things
do we have to do on a schedule?
Especially at work, it's like make sure
you run this report every Monday. Make
sure you answer this client and do so on
a bi-weekly basis. Make sure that you
understand how we make sure that you
summarize these three threads and put
them into a clear note for the VP by
this Friday afternoon. All of this stuff
is recurring and we know we have to do
it. When I was a marketing analyst, I
had to do weekly reports for my bosses
and I would have killed for this because
so much of it was basically going around
the web and pulling this from this
logged in state in Marquetto and pulling
this from this logged in state in
Instagram and this and that and the
other thing. It's so nice to just have a
workflow that you can record and run and
then it gets the data. Basically, if you
can record something and get Claude to
do it for you, the world becomes your
oyster. Let's jump to another one that's
going to be real popular. Inbox triage.
So, this is one that's popular because
so few people like doing email. In fact,
if you look at the most popular use case
for OpenClaw, it's just doing email. But
you don't have to go and use OpenClaw
for that. You can open up Gmail in your
browser and Claude as an extension in
the browser as a little side panel
recognizes that you're there. It pops up
the Gmail icon and it says, "Hey, what
would you like to do?" You can have
Cloud scan your inbox, identify
marketing emails, identify newsletters,
pull up important messages. One of the
things I want to call out here, the
reason this is the third use case and
not just, hey, you can also do this is
that Anthropic recognizes that email's
really popular and Anthropic has done
work to make sure that Claude plays
nicely with the most popular email
service in the world, which is Gmail.
And so, support documentation confirms
that Claude has built-in knowledge of
how to navigate popular platforms like
Gmail. And so you don't have to give it
step-by-step instructions. In other
words, Anthropic is taking these popular
places where we spend most of our day on
the web, and it's making sure Claude
knows and recognizes how to use them
from the get-go without specific
instructions from you. And that is a big
deal. But you don't have to just take my
word for it. Techraar's Eric Schwarz
also tested this in his hands-on review
of the Claude extension. He tested both
the email side and the calendar
management side. And he really got it to
do meaningful work. Claude scanned his
Google calendar. It proposed open time
slots. It drafted an email to guests for
an event. And he could get Claude to
organize roughly 900 loose documents in
Google Drive because, oh, that's right.
Claude also recognizes Google Drive.
Claude was able to create a logical
folder structure for him. It was able to
sort documents into subfolders. It was
able to flag duplicates. So Claude was
able to basically work within the Google
ecosystem for email, for calendar, for
Docs, for Drive and put it together in a
way that allowed Eric to do useful work
at scale. Now, one thing I want to
caution about here is doing automated
email replies. Emails are high value,
especially emails to important
stakeholders. What I would not do at
this stage is tell Claude, "Please find
the important emails in my inbox and
autodraft replies." To me, that runs the
risk that Claude will send the wrong
message to the wrong person,
accidentally hit send instead of saving
draft. Basically, use this for inbox
cleanup. Use it for Google Drive
cleanup. Use it for calendaring. That
all works well. Be cautious about the
send function and the draft function
until you are sure that you are not
going to risk sending to a stakeholder
without you putting eyes on it. But of
course, sometimes we have work that
needs doing over multiple tabs. we have
to put together multiple recipes into a
single dinner for tonight or we have to
look at multiple competitor websites and
grab pricing from all of them. Instead
of having Claude do that work in an
extension site by site, you can save
time by having all of the sites pulled
up and having Claude tackle them in a
group. And there are two big ways to do
this. You can either do this with the
Claude extension in Chrome, which is
most of what we're talking about today.
It's worth noting that you can do this
from co-work as well because co-work can
navigate to a Chrome group tab just like
the claude extension. If you're thinking
there's a pattern here and Enthropic is
giving us lots of ways to do the same
thing, you would be right. Enthropic
wants you to work where you're
comfortable and wants to give you basic
tools that you can use to do a lot of
useful work. And the mechanics are
pretty simple. Claude can retrieve data
from all of the tabs in the group that
it has permission to see. And that's
designated by a clear group tab in
Chrome. Chrome launched group tabs
already. All you're doing is saying in
this group tab, Claude is active. And so
anything inside that group tab, Claude
can see and Claude can't see anything
outside that group tab. So just drag
tabs into Claude's designated tab group.
It views and interacts with all of them
at once. You don't need to switch
between tabs. You don't need to switch
so that Claude can see things. Claude
can automatically read across all those
group tabs in Chrome, synthesize the
content, and produce a structured
output. For example, if you have a
potato recipe and a chicken recipe up,
it can put together a potato and chicken
dinner recipe. It can put together a
meal plan for a roast potato and chicken
dinner and give you a complete
ingredients list and a complete plan for
cooking and having fun with board game
night. The nice thing is you can combine
this with some of the other stuff. And
so if you have, say, a competitor
pricing thing you need to run, instead
of showing Claude how to do that by
opening the tab one at a time, you can
have all three tabs up at once in the
group, you can show Claude what you want
to do on each tab, and Tab can run that
workflow, pulling data from those tabs
simultaneously. I'm picking these
examples because they give you a sense
of the elemental building blocks of this
tool and they help you to be able to do
useful work. Now, let's say you want to
go a step further. you want to get to a
full Excel file instead of just getting
to information that you can print in a
chat window. That's super easy. But in
this case, you're going to have to use
the co-work approach to working with
tabs. So remember, I said there's two
ways to do this. You can do this in
claude the extension or you can do it in
co-work
from enthropic. Well, in co-work all you
have to do is say work with this group
of tabs which I just described and then
you say in addition once you pull the
data out like all the competitor pricing
information please make it into an Excel
file with this format and it will. And
so this gives you the ability to do real
meaningful work that involves extracting
data from the internet pulling it from
multiple tabs putting it into a
structured output getting a specific
document or spreadsheet or PowerPoint
presentation out at the end. Basically,
stringing together these seemingly
simple things like looking at multiple
tabs becomes a way to get a lot of
useful work done from your agent while
you do other stuff. Now, this one is
specifically for developers, but I think
it's a really, really big deal. Now,
this is for anyone who has ever run
across a bug on their site that they
want to fix. One of the nice things
about essentially giving Claude eyes
inside the Chrome browser is that you
have eyes on the most popular browser on
the planet which renders so much of the
internet that we all see and consume
every day. And so if you're building
something you want the rest of the
internet to see and you want to test it,
you obviously have to test it in Chrome.
And now Claude can see it directly. And
oh by the way, if you are building with
Claude code, it gets even easier because
yes, you can have Claude test your
website on a schedule and look for bugs.
That's something you can set up as a
recording in Claude's extension to do.
And so if you're building something for
the web, you can actually give Claude
the eyes to see that either through the
Claude code terminal if you're someone
who's a little bit more technically
inclined or you can do some version of
this through the extension as well
because remember you still have the
ability to record workflows. you have
the ability to schedule things in tabs,
etc. And so what you can do is you can
go to a website you're trying to work on
and you can record a specific flow using
the extension like you can do a test
checkout and you can say please run this
test checkout every Thursday or every
day or every morning at 9:00 or whatever
it is and Claud's the extension will run
that and it will be a very easy way to
test if your product is still up and
running. Developers have faster ways to
do that. Right? I'm saying if you're a
non-developer this is a super easy way
to do it. If you're a developer and
you're in Cloud Code already, you can
just use Cloud Code to navigate and hit
the browser and actually immediately
test aspects of your product. And yes,
that includes these periodic what
developers call smoke tests where you're
testing if the checkout works. Basically
includes lots of ways to debug. And as
funny as this sounds like, it sounds
like I'm just saying you can do really
basic debugging and if your site works,
why do you care? But you can actually do
much more sophisticated things if you're
willing to go to the terminal, which
again, don't be scared of the terminal.
It's not going to bite you. For example,
you can take a Figma mock. You can have
Claude build it. You can have Claude
verify the accuracy of what Claude built
in the browser while looking at the
Figma mock. And then you can have Claude
debug it live. Essentially, if you give
Claude eyeballs and you enable Claude to
build, Claude can build something it can
see and check. And so the experience is
like watching code that you care about
for customers being written for you in
your terminal while Chrome autonomously
opens tabs and clicks through your UI
and reads console output and bugs and
reports back to the claude code in the
terminal and the agent works in the
terminal to correct it. So one agent is
writing the code. Another version of
Claude is testing in the browser. The
human is just watching. And this whole
loop used to require me as a product
manager sitting with a developer,
sitting with a QA engineer, sitting and
understanding how this works in a
staging environment. Now this just runs
on local in a loop until it's done. Now
I understand that we have larger tasks
we need to do in staging environments
and enterprise software is not always
this simple but these basics matter
because if you're trying to build
something that is useful you always
start with the idea of what does the
customer see and that's true whether
you're at an enterprise or a small
medium business and the ability to get
to that very quickly to see if it
matches the mock to see if you have the
ability to render a component correctly
to see if the input field works those
are things that we have spent an
inordinate amount of time validating for
anyone who works in software. Like
catching bugs is a painful exercise.
Giving Claude eyes gives Claude the
ability to build with much fewer errors
than otherwise. Now, you might think
this is just a pumpup post for the
Claude extension, blah blah blah. I'm
going to be honest about some of the
limitations to it. I'm going to use a
specific example. I validated this, but
this is also something that Neuron
called out in January, and it's
something I expect to get fixed as
models get better, but it is a
limitation of the Claude extension
today. If you give Claude a dataheavy
task, Claude has limitations in Chrome
on how much data it's actually going to
manage. And so the Neuron example is
LinkedIn contacts. If you want to record
a weekly workflow where Claude will scan
LinkedIn pages, we'll review content
across posts, we'll interpret contacts,
we'll provide summaries, that all
technically works. But what Neon found
is if you expand that watch list beyond
a few people posting, coverage can get
spotty. Expected posts sometimes don't
surface. One summary might focus on a
tangential update that might not be
super relevant. Essentially when you
start to expand the scope of the task in
a single workflow right now claude in
Chrome does not do a perfect job of
recognizing salience of recognizing
what's really important and it doesn't
do a perfect job of coordinating and
synthesizing all of that data in a
useful manner. It can miss stuff. Now, I
do want to call out that I expect that
particular aspect to get better
relatively quickly as models continue to
improve. But fundamentally, it remains a
challenge because what you're doing is
you're feeding this LLM all of this open
context from the web and the LLM has to
look through that context window and
find what really matters. And that can
be hard. And so if you're giving Claude
a data heavy task in Chrome, I would
recommend breaking it up into subtasks
because I think you're more likely to
get useful results done, especially if
this is a recorded workflow and you want
to run it on a schedule. You don't want
mistakes. You'd rather do a very clean
subtask. So if you ladder all of this
together, what does it mean? The pattern
across all of these use cases is
similar. It's not the AI assistant
answers questions while I browse, which
is what people often think of with these
things. It's actually the browser agent
does real work on my behalf. It clicks,
it navigates, it reads, it extracts, it
does all of the things I do on the web
to generate value, but it does it
without me getting involved, especially
if it's on the schedule. And that
distinction really matters because it
changes what you're optimizing for. With
a chatbot, you optimize for your own
questions, right? With a browser agent,
you optimize for your workflows. So, the
skill isn't prompting. The skill is
looking at the repetitive work you do
every week and asking yourself, can I
describe this clearly enough that an
agent can do it for me on a schedule
without supervision? And I think it's
worth calling out that that skill is
something that generalizes in the AI age
because so much of what we are going to
be doing in 2026, whether you're working
with Claude or Chat GPT or another LLM,
it's going to involve giving the LLM
context on a repetitive piece of work
you're doing and asking it to do it. And
the UI can look different. Sometimes
you're in the Chrome tab, sometimes
you're in Cloud Code, sometimes you're
in co-work, sometimes you're in chat
GPT, which also has scheduled tasks
directly. Well, the principle is the
same. You have to understand what you
want to do and be able to either mimic
it and do it directly, which is really
convenient in Chrome with that record
button from Claude, or you have to
describe it clearly enough that it can
be done anyway by an LLM, which is what
you would do with a recurring task in
chat GPT. Now, I do want to call out
that anytime you put an LLM into the
open web, there are nonzero risks. You
should be responsible and you should use
Claude on trusted sites. Do not go to
weird corners of the internet with this
tool and expect not to get prompt
injected. There are ways for open text
on the internet to hijack your agent and
give it malicious instructions. Like for
example, if you went to a weird corner
of the internet like corner of Reddit
with a prompt injected thread in the
page and you also have your email up in
the same tab, then theoretically your
LLM could get prompt injected simply by
reading the Reddit thread in the page
and go to your email and start sending
sensitive data. That is a real example.
And so use this tool on trusted sites.
Please review sensitive actions. I
suggested replies to stakeholders
something that's sensitive. you will
have others. Don't open your bank
account on this site. Like, don't be
dumb. Treat this like a capable but new
employee. Verify the output. Don't
assume that this thing has infinite
permissions. Now, this extension is
available to anybody that is on a paid
plan right now. The degree to which you
can pick a model for intelligent tasks
will depend on what kind of plan you're
on. In general, the simpler your plan,
the dumber the model that's available.
And dumber models have difficulty with
ambiguity, with data. And so for
instance, my example with LinkedIn data,
you may want to have something like a
max plan or a team plan or an enterprise
plan to do a complex task like that
where you tell the browser extension to
go to multiple LinkedIn profiles and to
summarize what they posted and put that
back to you. That's like a bunch of
browsing. It's a lot of context window
to manage, etc. And if you get a basic
pro plan, the model may not be smart
enough to do that well. Ultimately, when
we were in early 2025, mid 2025, even
fall of 2025, some of the question was,
"Does this work? Can the agent actually
navigate the internet?" We are past that
now. In early 2026, the question is not,
"Does this work anymore?" It does work.
People are saving dozens of hours a
week. They're getting rid of a lot of
repetitive work. The question for you
is, do you understand the tools
involved, which I spent a lot of this
video on, and can you identify the
repetitive work clearly enough so that
you can go and offload it onto tools
like the anthropic extension for Claude
inside Chrome. So there you go. I hope
this has given you a sense of why this
matters, why this is not just a chatbot
inside a browser, and you have a sense
of the useful work that you can do. If
you want a bunch more examples, I put
them all in the substack and I have a
whole guide there.