Claude in ChromeHow an agent works in your browser — tests, debugging and forms under your login
Claude in Chrome connects Claude Code to the browser: the agent opens tabs, reads the console, runs scenarios and fills forms under your login. How to connect it, what it does and which permissions to give.
Claude in Chrome connects the agent to your browser: it opens tabs, reads pages and the console, clicks and types under your login. Below: how it works, how to connect it in a few minutes, what the agent can do and which permissions to give. The agent from the QA with an agent article uses this channel, and connecting works like MCP.
What Claude in Chrome is
A Chrome extension connects Claude Code to the browser. The agent opens new tabs and works in a visible window.
ChromeEdgeBraveArcVivaldiOperaWSL: no
Takeaway
In this exampleThe agent in the terminal drives Chrome tabs through the extension and uses your logins.
In generalActions are visible in the window in real time. On a login page or CAPTCHA the agent stops and waits for you. In Claude Code the extension appears as the claude-in-chrome MCP server: /mcp lists its tools.
What to doCheck the requirements: extension version 1.0.36 or later, Claude Code, a Pro, Max, Team or Enterprise plan, and sign-in with /login.
How to connect
Tell the agent
Open code.claude.com/docs, find the search box, type “hooks” and tell me which results appear.
In this exampleThe agent opens the docs, types a query into the search box and retells the results. That checks the connection works.
In generalConnecting takes minutes: the extension comes from the Chrome Web Store, and Claude Code does the rest.
What to doRun the example and make sure /chrome shows “Status: Enabled” and “Extension: Installed”.
What the agent can do in the browser
Takeaway
In this exampleOne connection covers debugging, tests, forms, documents and data extraction.
In generalThe agent reads the console and DOM, runs scenarios, fills forms, uploads files (up to 10 MB in total) and records GIFs.
What to doPick one repetitive browser task and hand it to the agent.
Example: a bug on the test stand
Ticket PAY-142 from the article on QA with an agent: the agent walks the path itself and brings evidence.
Tell the agent
Open the test stand, add an item to the cart, apply the promo code and pay with a test card. If the payment fails, look at the console and record a GIF of the path.
Takeaway
In this exampleThe agent walks the path, finds the console error and records a GIF for the ticket.
In generalThe window is visible, so QA follows the steps and can stop the agent. Console text and the GIF leave a checkable trail.
What to doAsk for a GIF and a console excerpt with every result.
Permissions and risks
SitesThe extension settings control which sites the agent can browse, click and type on.
Your loginsThe agent can reach everything you are signed into in this browser.
Login and CAPTCHAThe agent pauses and waits for you.
GIFThe recording includes everything visible, account details too: review it before sharing.
Dialogsalert, confirm and prompt dialogs block commands: close them by hand and ask the agent to continue.
Page text can be aimed at the agent. This risk is covered in a separate article.
Takeaway
In this exampleAllowed sites are open to the agent, the rest stay closed, and actions on a new site need your consent.
In generalThe agent works with your logins, so site permissions are the first line of defence.
What to doOpen the extension settings and keep only the sites the task needs.
When you need your own tool
Claude in Chrome suits one-off work, but production tasks often need a tool of your own.
PuppeteerA library that drives the browser from code: open a page, click, pull out what you need. It runs without an agent at every step.
Whole pageClaude Code takes the whole page: text, elements, screenshots. That is convenient for a one-off task.
Tokens and speedThe bigger the page, the more tokens each step costs: slower and more expensive.
Your own scriptA custom solution takes only the part of a site or page you need: a table, a form, one block. The agent gets a short result.
What to chooseOne-off tasks and exploration: Claude in Chrome. Regular, high-volume work with a token budget: your own script the agent calls as a tool.
The script can be connected to the agent through MCP or a skill: the agent asks for the piece of the page it needs instead of opening all of it.
Takeaway
In this exampleThe agent reads the whole page and spends tokens and time on it.
In generalFor production, tokens and speed matter, so only the needed parts of sites and pages are taken.
What to doLeave one-off work to Claude in Chrome. Move repeating work into a Puppeteer script and give it to the agent.
Where to start
Takeaway
In this exampleThree steps: extension, launch, check.
In generalStart with a local stand: a mistake there breaks nothing.
What to doToday, install the extension and run the docs example.