Browser Control
Most of what a coding agent does needs no browser at all. This is for the part that does - the cloud console, the dashboard, the admin panel, and everything else that has no API to call and no way in from the outside.
TheGitAI is an agentic AI coding tool for your terminal. With the **TheGitAI Chrome extension** installed, it can also work in the Chrome you already have open, signed in as you already are. There is no second browser, no separate sign-in, and no API key standing in for you.
What this is not
It is not mainly a front-end debugger.
Checking how a page you just built renders is the one browser task that needs nothing special: the page is local, it has no session, and Playwright or a headless script does it perfectly well. If that is the whole job, use one of those.
What a headless script cannot do is work in a browser that is already yours. No cookie jar to transplant, no credentials to hand to a tool, no screenshots to paste in so the agent can see what you are looking at. That gap is the reason this extension exists.
Install
Two halves, and both are required. The extension does nothing on its own.
1. **The terminal app**, if you do not already have it:``` npm install -g @thegitai/cli ``` 2. **The extension**, from the Chrome Web Store.
Then run `ai` in any project directory and ask about a page. The extension connects on its own - there is no pairing step, no port to configure, and nothing to sign into twice.
Chrome 116 or later is required.
What to ask for
- **Work a console owns.** Deploy a model, provision a resource, or change a setting through a cloud provider's portal, when the UI is the interface and working out the API would take longer than the task.
- **Data out of a site with no export.** No download button, no API, but it is right there on the page in front of you.
- **Admin UIs.** A CMS, a page builder, an internal dashboard - anywhere the job is clicking through screens rather than writing files.
- **A walkthrough on your own screen.** Ask where a setting lives in a panel you do not know, and have it shown to you a step at a time.
- **The long form** on the internal tool, the one you have already filled in twice today.
- **What is on a page you are signed in to.** Your own profile on a professional network, a feed, a job posting - read it and do something useful with what is there.
- **Front-end work too, when you want it.** How a page renders, what breaks at phone width, the console error behind a failure. Real, just not the reason the extension is here.
Its clicks and keystrokes are the same kind the browser gets from you, rather than scripted events - which is why pages that ignore automated input still behave normally.
Staying in control
| Boundary | How it works | | --- | --- | | Tab group | Each terminal session owns one Chrome tab group. A tab joins when you ask the agent to use it; dragging it out takes it back out of reach immediately. | | Your other tabs | Never read. The agent is told how many you have open - never their titles, never their addresses. | | Approval | Default mode asks before the agent uses a website. Plan mode lets it read pages without clicking, typing, or navigating. | | Downloads | Paused the moment they start. The filename, size, and origin are surfaced, and nothing reaches your disk until you accept. |
| Chrome profiles | With more than one Chrome profile connected, TheGitAI asks you which one to use the first time it acts in the browser, listing them by the names they have in Chrome, such as "Work (Profile 3)", with the one you used last as the default. It stays in that profile for the conversation, including after you resume it, and asks again before switching. | | Stop | **Stop**, **Resume**, and **Disable** live in the toolbar popup. Chrome also shows its own banner for as long as the agent is attached to a tab. | | Credentials | It never types your passwords. Logins and CAPTCHAs are handed back to you. |
Exiting the CLI leaves your pages open. Another session cannot take tabs owned by a live one.
The same session that makes this useful is the one that makes it worth watching. A page can carry instructions written to be read by an agent rather than by you. Default mode asks before the agent uses a site for exactly this reason - start on sites you know, and read what it tells you it is about to do.
Privacy
The extension makes no network requests of its own. Everything it reads goes to the TheGitAI tool on your own computer, over a local channel between two programs on the same machine.
Nothing it reads is stored by us - not the pages, not the screenshots, not a record of the sites you visit. On your own machine it keeps exactly two things: a random identifier so two Chrome profiles can be told apart, and whether you have switched it off. Uninstalling removes both.
See Trust & Privacy and the full privacy policy.
When it does not connect
The popup tells you which half is missing. If it says TheGitAI is not installed on this computer, the terminal app is not there yet or has never been run - install it and run `ai` once. The CLI registers the local connection on every interactive launch, so it repairs itself.
When a browser request starts without a connection, TheGitAI waits up to 2.5 seconds for discovery before reporting a failure. This includes the first browser request in a session as well as reconnects. A failed check does not mean the extension is missing or disabled, and a later request checks again. Open the TheGitAI icon in Chrome to check its connection status if the browser remains unavailable.
After updating the CLI, exit and start a new `ai` process to load the fix. Existing processes continue using the code they started with. If a connection is replaced while an action is running, TheGitAI reports that its outcome is unknown. Check the page before repeating the action; it is not replayed automatically.
See Troubleshooting if it persists.
Return to a normal website tab before asking. Chrome does not allow any extension to control its own settings or extension pages, so a request made while `chrome://` is in front will have nothing to act on.