Think about where your genuinely tedious work actually happens. Not in your documents — in a browser tab. The CRM you paste contacts into one at a time. The supplier portal with the form that never remembers anything. The five internal tools that don’t talk to each other, so you’re the integration.
An AI that can only read your files is useless for all of that. This is the feature that isn’t.
What it does
Claude can drive a real browser: open tabs, navigate, click, type, fill forms, read what’s on the page, upload files, and pull structured data out of sites. Crucially, it shares your existing login state — so it works inside the tools you’re already signed into, with no API keys, no integration project, and no help from IT.
That last point is the one that changes what’s possible for a non-technical team. “Can we connect Claude to our CRM?” has historically meant a conversation about APIs and permissions and a quarter of engineering time. If Claude can simply use the CRM the way you use it, that conversation stops being necessary.
Two ways to get there
On Desktop, there’s a browser built in. The Claude desktop app includes its own browser, so Claude can pull up a site, read documentation, look at a design, or interact with a page without you leaving the app. Nothing to install; it’s part of the window you’re already in. For most non-technical people, this is the whole story — start here.
In the terminal, it connects to your Chrome. The Power-Track route uses the Claude in Chrome extension to drive your actual browser from the CLI or the VS Code extension. You install the extension from the Chrome Web Store, start Claude with claude --chrome, and approve the connection. Run /chrome at any time to check status, reconnect, or set it to be enabled by default. This is the path to reach for when you want browser work chained into scripted or automated workflows.
Both are the same underlying capability. The Desktop version removes the setup; the CLI version adds automation.
What it’s genuinely good at
The examples worth caring about are the boring ones.
Repetitive data entry. You have a spreadsheet of contacts and a web form that takes them one at a time. Claude reads your local file, opens the site, and works through the rows:
I have a spreadsheet of customer contacts in
contacts.csv. For each row, go to the CRM, click “Add Contact”, and fill in the name, email, and phone fields.
This is the single highest-value use for most business teams, and it’s the one nobody demos because it isn’t impressive. It just removes an afternoon.
Getting data out of sites that won’t export it. Plenty of internal tools and public pages hold information in a shape you can’t download. Claude can read the page and produce a structured file — “extract the name, price, and availability for each item and save it as a CSV.”
Working inside apps you’re logged into. Because it inherits your session, Claude can draft directly into a Google Doc, work in Gmail, or update something in Notion — without any connector setup. For someone whose company hasn’t wired up integrations, this is the shortcut around the whole problem.
Multi-site workflows. The chains that make up half of coordination work: check the calendar for tomorrow’s meetings, look up each external attendee’s company, and note what they do. Three tools, one request.
Checking your own work. For anyone building something, Claude can open the page it just changed, click through it, read the console for errors, and fix what it finds — closing the loop without you playing tester.
Recording what happened. It can capture a browser session as a GIF, which is a quietly useful way to document a process for a colleague. Worth knowing: the recording captures everything visible, including account details on logged-in pages, so review before sharing outside your team.
How much control you keep
This is the feature where people’s instincts get nervous, and mostly for good reasons. So, specifically:
You watch it work. Browser actions run in a visible window in real time. You see the tab open, the field fill, the button get clicked. This is not something happening invisibly in the background.
It stops at the hard parts. When Claude reaches a login page or a CAPTCHA, it pauses and hands control back to you. It doesn’t attempt to authenticate as you.
Permissions are per-site. Site-level access is inherited from the Chrome extension’s own settings, so you decide which sites Claude may browse, click, and type on — and you can change that at any time in the extension.
Reading and acting are treated differently. In plan mode, calls that only read the page — reading text, checking the console, taking a screenshot — run without a prompt, while anything that changes state — clicking, typing, navigating, managing tabs — asks for approval first. That’s a sensible default: looking is free, doing needs a nod.
The risk worth taking seriously
Here’s the honest version, because it deserves more than a footnote.
You are giving an AI the ability to act inside sites where you are already authenticated. That’s exactly what makes it useful, and exactly where the risk lives. A web page can contain text crafted to influence a model reading it — and unlike a document you chose to open, you don’t always control what’s on a page.
Three habits handle nearly all of it:
- Be deliberate about site permissions. Grant the tools you actually want automated. Don’t grant everything because it’s convenient.
- Keep the approval gate on for anything that sends, posts, or submits. Reading a page is low-stakes; submitting a form on your behalf is not.
- Don’t mix untrusted pages with sensitive access in the same session. If Claude is reading arbitrary sites, that’s not the session in which it should also have your finance portal open.
None of this makes it unsafe to use. It makes it a tool that deserves the same care you’d apply to giving a new contractor your logins — which is to say, some.
Where it fits alongside everything else
There are now three ways Claude reaches your work, and it’s worth knowing which to reach for:
- Cowork — for files sitting in a folder on your machine.
- MCP and connectors — for proper, sanctioned integrations with systems that have an API.
- Browser control — for everything else: the tools with no API, the internal portal nobody will integrate, the site you just need something done in.
Browser control is the pragmatic one. It’s not the most robust route — a redesigned page can break a workflow that an API wouldn’t — but it works today, on the tools you actually have, without asking anyone’s permission.
Try it with something small
Pick the most irritating repetitive thing you do in a browser. Not the important one — the annoying one. Forty rows into a form, a weekly copy-paste between two systems, pulling a list off a page that won’t export.
Describe it, watch the tabs open, and check the result. That first watch is the part people remember: it’s the moment the abstraction “AI agent” turns into something with hands, doing a job you were dreading.