Your agents drive
your real Chrome
Install the AgentsRoom extension, switch on "Let my agents drive this browser", give the browser a name. From then on any agent can open a tab in that Chrome, with your sessions, your cookies and your extensions, and click, type, read the console and take screenshots there.
One Chrome profile is one named target. So one agent can work signed in as your pro account while another works signed in as your personal one, and the Chrome can sit on another machine of yours.
browser_select_target "Chrome · Pro"browser_select_target "Chrome · Personal"browser_navigatebrowser_clickbrowser_screenshotTwo agents, two Chrome profiles, one encrypted relay in between. Each agent picks its target by name.
AgentsRoom already ships an embedded browser inside every project, and that stays the default: it is where an agent tests your localhost without touching anything of yours. But some work only makes sense in your actual browser. Checking a deploy on an admin panel that requires your SSO, filing a form on a service where only you are signed in, reproducing a bug that only shows up with your account and your extensions. That is what Chrome control is for.
The mechanism is the AgentsRoom Chrome extension. When you switch driving on, the extension announces the browser to your account under the name you gave it, and the desktop lists it next to the embedded screen. The agent gets the same MCP tools it already has for the embedded browser, plus one to pick the target. Nothing else changes for the agent, which is the point.
Chrome shows its own banner, "AgentsRoom started debugging this browser", for as long as an agent drives. It is the same banner every debugging extension triggers, and it is your guarantee that nothing drives your browser silently.
One Chrome per account
Chrome runs one instance of an extension per profile. AgentsRoom turns that into a feature: each profile registers as its own named browser, with its own cookies and its own signed-in accounts. You name them the way you think about them, Pro, Personal, Client, and the agent asks for one by name.
Since every agent has its own target, several agents can work on several accounts at the same time. A QA agent verifies the staging site in the profile signed in as the test user while a second agent updates a ticket in the profile signed in as you, and neither sees the other's session.
One profile, one target
A Chrome profile is a set of cookies, sessions and extensions. Each one shows up as a separate browser the agent can choose, so a signed-in session is never shared by accident.
Named by you
Chrome does not expose the profile name to an extension, so you type one in the popup, with suggestions like Pro or Personal. That name is what the agent uses to pick the target.
Several agents at once
Targets are chosen per agent, not per project. Two agents on two accounts, or five agents on the same Chrome, each in its own tab, run side by side.
Your browser, on any machine
The extension does not talk to the desktop on localhost. It talks to the AgentsRoom relay, the same one the mobile app uses to reach your desktop. So an agent running on the Mac at the office can drive the Chrome on the PC at home, as long as both are signed in to your account and driving is switched on there.
The relay sees encrypted blobs and nothing else. Every message, every screenshot, is sealed with a key derived from your account and shared only by your devices, so the relay cannot read a URL, a page or a keystroke. The same design, and the same guarantee, as the mobile companion.
Through the relay, not localhost
A browser announces itself to your account every 30 seconds. The desktop lists the ones that are online and greys out the ones that stopped answering.
End-to-end encrypted
Ops and results travel in a dedicated session room between one desktop and one browser, sealed with a key the relay never holds.
Same principle as mobile
If you already pilot a desktop agent from your phone, you are using this exact transport. Chrome control adds a browser to the set of things it carries.
Every agent, every CLI
Chrome control is an MCP server, the same AgentsRoom Browser MCP that drives the embedded browser. That is why it is not tied to one model or one vendor: Claude Code, Codex, GitHub Copilot CLI, Cursor and 10 other agent CLIs get the same tools, with the same names, and pick their target the same way.
Concretely the agent calls browser_list_targets to see which browsers are online, then browser_select_target with a name or an id. From there browser_navigate, browser_click, browser_type, browser_screenshot, browser_evaluate and the rest act on that Chrome instead of the embedded screen, until the agent switches back.
That includes the CLIs that ship no browser extension of their own. OpenCode, Grok Build, Mistral Vibe, Kimi Code or Aider get a real Chrome, with your accounts in it, the day you install the AgentsRoom extension: the extension belongs to AgentsRoom, not to the model vendor, so no CLI has to build one.
browser_list_targetsbrowser_select_targetHow it works
Install the AgentsRoom extension
From the Chrome Web Store, in each Chrome profile you want to make available. Sign in to the same AgentsRoom account as your desktop.
Switch on "Let my agents drive this browser"
In the extension popup. Chrome asks for the debugger permission at that moment, not at install time, and the extension only attaches to a tab while an agent is actually driving.
Name the browser
Pro, Personal, Client, whatever tells you which account is signed in. The name is saved as you type and announced to your desktop immediately.
The agent picks its target
Either the agent calls browser_select_target with the name, or you pick the target for it in the Browser panel of the desktop, next to the embedded screen.
The agent drives, you watch
The first action opens a fresh tab so the agent never takes over the one you are reading. Chrome shows its debugging banner for the whole session. Switching driving off closes the sessions and detaches the debugger.
The same extension already does four other things
Chrome control is the newest tab of the AgentsRoom extension, not a second extension. If you already use it to file tickets or to point at an element, you have it: same popup, same account, one more switch.
Point on page
Click any element on a page and it lands in the composer of your active agent with its HTML and a selector. The fastest way to say "this button, right here" to an agent.
Learn moreSend page to my desktop
One click drops the HTML of the current page into your active agent's message box, over the local loopback, never through a server.
File a ticket from any page
Title, description, screenshot, straight into the backlog of the project you pick. A bug seen on staging becomes a ticket an agent can start on.
Learn moreProject picked by URL
Link URL patterns to a project and the extension preselects it when you are on that site. No dropdown to open when you file from the app you are working on.
Claude in Chrome alternative
An alternative to Claude in Chrome, for every agent CLI
Claude in Chrome is Anthropic's extension: it puts Claude inside the Chrome profile it is installed in, on the machine it runs on. If you work with Claude alone, on one account, on one computer, it does the job and you need nothing else.
AgentsRoom Chrome control is the alternative for the cases it leaves out. The extension belongs to AgentsRoom, not to a model vendor, so the same Chrome answers Codex, GitHub Copilot CLI, Cursor, OpenCode, Grok Build or Aider as well as Claude Code, and a CLI that ships no browser extension of its own gets one this way. Each Chrome profile is a named target, so several agents work on several accounts at the same time, and that Chrome can sit on another machine of yours.
When you want the alternative
Your agents are not Claude
They run on Codex, GitHub Copilot CLI, Cursor, OpenCode, Grok Build, Mistral Vibe, Kimi Code or Aider. Claude in Chrome answers none of them. The AgentsRoom extension answers all of them, through the same MCP tools.
You need more than one account
Pro, personal, client: every Chrome profile is a target an agent picks by name, so two agents can be signed in as two people at the same time, instead of sharing the one profile the extension lives in.
The browser is somewhere else
The Chrome that holds your sessions is on your desk and you are on the laptop, or the other way round. The extension talks to the AgentsRoom relay, end-to-end encrypted, so an agent drives it from wherever it runs.
You want the browser inside the orchestrator
Chrome control is one tool among the others in AgentsRoom: backlog, dev terminals, embedded browser, QA agents. The browser is one step of the agent's work, next to its terminal, its tickets and its tests.
What you keep: a real Chrome, your sessions, your cookies, your extensions, and Chrome's own debugging banner so you always know an agent is driving. What changes: who can drive, from where, and on how many accounts.
AgentsRoom vs Claude in Chrome
Both let an AI agent act inside a real Chrome. The difference is in who can drive, from where, and on how many accounts.
| Capability | AgentsRoom Chrome control | Claude in Chrome |
|---|---|---|
| Accounts an agent can drive | Every Chrome profile, picked by name | The profile it runs in |
| Several agents on several accounts at once | Yes, one target per agent | Not documented |
| Browser on another machine | Yes, through the encrypted relay | Same machine |
| Agent CLIs | Claude, Codex, GitHub Copilot CLI, Cursor and 10 other agent CLIs, including those with no browser extension of their own | Claude |
| End-to-end encrypted transport | Yes, key shared by your devices only | Not applicable, stays on the machine |
| Embedded browser as well | Yes, and it stays the default | No, it is the Chrome itself |
The Claude in Chrome column reflects Anthropic's public documentation at the time of writing. Check their current documentation before relying on a cell, the product moves.
FAQ
Do I need the desktop app, or is the extension enough?
You need both. The agents run in the AgentsRoom desktop, the extension is what makes a Chrome reachable to them. An agent started from the mobile app gets the same tools, since it runs on the desktop anyway.
Can an agent take over the tab I am reading?
No. The first action opens a new tab, and the agent only switches to an existing tab when it explicitly asks for it. Chrome also shows its "started debugging this browser" banner for as long as an agent drives, so you always know.
Which agents can use it?
All of them. Chrome control is part of the AgentsRoom Browser MCP server, so Claude Code, Codex, GitHub Copilot CLI, Cursor and 10 other agent CLIs see the same browser_list_targets and browser_select_target tools. A CLI that has no browser extension of its own, OpenCode, Grok Build or Aider for instance, gets one this way.
What does the relay see?
Encrypted blobs. Every message and every screenshot is sealed with a key derived from your account and known only to your devices. The relay routes the bytes, it cannot read a URL, a page or what the agent typed.
Does this replace the embedded browser?
No. The embedded browser stays the default target, and it is still the right place to test your localhost without touching your sessions. Chrome is an extra target, chosen per agent, for the work that needs your real accounts.
How do I stop it?
Switch off "Let my agents drive this browser" in the extension popup. The browser tells your desktops goodbye, the sessions close and the debugger is detached. Agents pointed at that Chrome fall back to the embedded screen.
Is AgentsRoom an alternative to Claude in Chrome?
For three cases, yes: your agents are not Claude, you need several accounts driven at the same time, or the Chrome is on another machine. If you use Claude Code alone, on one profile, on one computer, Claude in Chrome already covers you. The two extensions can be installed side by side: Chrome control only drives when you switch it on, in tabs it opens itself.
You may also like
Browser automation
The embedded browser and the MCP tools an agent uses to test your localhost, with the same tool names Chrome control reuses.
Remote fleet
Run agents across every machine of your account from one window. Chrome control rides on the same relay.
AgentsRoom MCP
The MCP servers that let any agent CLI drive AgentsRoom itself: backlog, dev commands, prompt library, browser.
Every agent CLI
Claude, Codex, GitHub Copilot CLI, Cursor and 10 other agent CLIs, side by side in the same projects.
Give your agents your browser, one account at a time
Install the extension, name your Chrome, and let an agent work where you are already signed in.
Companion app: monitor your agents on the go
Bring your own: Claude, Codex, Antigravity CLI, or other AI provider.
Push bugs and requests straight to your public backlog.