Where a cloud browser agent works

Many AI browser agents run their browser on the vendor's servers. You watch a stream of it, and when a task reaches a page that needs your account, you sign in again, inside that remote browser. OpenAI's documentation for ChatGPT's cloud browser describes the model well: you type your credentials into the login screen it shows you, the signed-in session is kept for later tasks until it expires, and you can take over the remote browser when the agent gets stuck. The same page notes that some websites block automated browser agents even though they work normally in your own browser.

That design has real advantages: it runs while your laptop is closed, and it keeps the agent away from your own browser. The cost is that every account the agent touches has to be signed into a second time, on a computer you do not control, and some sites never let it in.

Where Computron works

Computron is a program on your computer. You type a request in your terminal (or in Chrome's side panel), and it does the work in a tab of the Chrome you already have open. The connection is a Chrome extension plus native messaging: Chrome starts a small helper on your machine, and Computron talks to it over a named pipe on Windows or a socket file on macOS and Linux. Neither is a network port. (The side panel is the one exception: it talks to Computron through a port on 127.0.0.1, which answers only requests that come from the extension.)

The task gets a tab of its own in a tab group named Computron, and the tab you are reading stays where it is. You can keep working while it runs, or turn on watch mode and see the cursor move. While a task runs, Chrome shows its "is debugging this browser" bar, and pressing Cancel on it ends Computron's access at once.

Your Google account comes with it

Because the work happens in your own Chrome profile, every site is already signed in as you: Gmail, Google Cloud Console, GitHub, your billing pages, the admin panel you open once a month. Nothing is copied to make that true. Your cookies, saved passwords, passkeys and extensions stay exactly where they are, in your profile.

Google sign-in is the common case where a site asks again. When a console you have not opened in a while sends you to Google's account list, Computron picks the account your request names, or the first one signed in, and presses Continue or Allow on the screens with nothing to fill in. It does this by matching names in code, without a model call. It never types a password. If Google asks you to confirm on your phone or with a security key, Computron moves the tab into a yellow Needs you group, tells you which tab it is, and carries on once you have tapped.

The same goes for command-line logins. gh auth login, gcloud auth login and similar tools finish in the browser, and the browser they need is the one where you are signed into GitHub or Google. Computron can approve those in your profile. Those runs are not recorded for replay, because the URL and code work only once, and sign-in codes are scrubbed from the logs.

What stays on your machine

Your sessions. Computron does not export cookies or copy your profile anywhere. The browser it drives is the one on your desk.

Passwords you type. When a page asks for a password Computron does not have, it asks you in the terminal, with the input hidden. The answer goes into the page's field, not to a model, and is saved in your operating system's keychain so you are not asked again. It is never written to one of Computron's files.

The records. Recorded tasks, run logs and the per-step decisions are plain JSONL files in a .computron folder, in the folder you start Computron from, plus settings in ~/.computron. Before one of them is written, Computron checks it against the secrets it knows about (values in your environment, and token shapes such as API keys), and refuses the whole write if one would leak. There is no delete command yet: to forget everything, delete those folders.

What leaves it

"Runs locally" does not mean "offline". The browser and the records are local; the thinking is not. Here is everything that goes out.

To the models that decide each step. On each step, Computron sends the page's accessibility tree, meaning the roles and names of the buttons, links and fields on screen, with known secrets replaced, to Jev, TypeSafe's decision model. When you ask for a value, up to 8,000 characters of the page's text go to Claude at the end of the run to read the answer off the page. The first time Computron meets a task, a Claude model works it out first in a background tab, reading page text and, when it needs to, a screenshot. A drag step sends one cropped screenshot around the handle to Jev. Your chat messages go to the chat model you chose.

To the shared map of websites. When a task succeeds, Computron learns which clicks lead from one page to the next, and it shares that with other Computron installs so a site you have never used before is faster the first time. What it sends: the site's host name in plain text (for example github.com), the structure of each page, which kind of action led where, and whether it worked. Page titles, button labels and paths are sent as hashes keyed to the site; the plain words are sent only after two different installs have sent the same hash, which means they are part of the site's own interface, not your project names. What it never sends: anything you typed, search terms, page text, or the address you started from. Each upload is signed with a key generated on your computer. This is on for every install today, and there is no switch to turn it off yet.

What local does not mean

It does not mean offline: without the model services, nothing runs. It does not mean background work while your computer sleeps: Chrome has to be open, because the work happens in it. And it runs in one Chrome profile, the one you loaded the extension into. If you keep work and personal accounts in separate profiles, load it into the one whose logins you want it to use.

Setting it up

Install Computron with the one command on the home page; the installer runs computron chrome install, which registers the connection and prints the one manual step: adding the extension from the Chrome Web Store. After that, tasks run in your normal Chrome with the logins you already have, and computron doctor tells you which connection is in use. If the extension is not available, Computron falls back to a separate Chrome started with a debugging port. That browser has its own, signed-out profile, so Computron tells you in one line when it falls back. The getting started guide has the full steps.