← back to the scrapbook issue #2026 · free · take one
← all field notes filed under: ari ✎

Ari can do what you'd sit down at a computer to do

#ari#ai#browser#computer#automation

Ask Ari for anything. If it can be done on a computer or online, Ari can do it for you.

That is a big sentence, so here is what it means and where it stops.

For months Ari could remind me, remember things, and search the web. That is useful, but it is still talking. Most of my actual to-do list is not talking. It is a PDF I need to fill out, a pile of photos in the wrong format, a booking form, an order I need to check on. The kind of thing where I sit down at a computer, click around for twenty minutes, and lose my afternoon.

So I gave Ari a computer and a browser.

Two things, not one

Ari’s computerAri’s browser
What it isA private cloud computer, one per personA real browser Ari drives on the live web
Good forFiles, scripts, installing tools, building thingsLogging in, filling forms, clicking through multi-step flows
You seeFiles in your Library, a “Working on” card while a job runsA live view of the page, in the chat
Ends withA file, a report, or an appA finished booking, a screenshot, a download

They work together. The browser downloads a statement, the computer pulls the numbers out of it. The computer builds the file, the browser uploads it to the site that wants it.

The computer

Ari’s computer is a Linux machine in the cloud. It is Ari’s, and it is yours, one per person. It is not your laptop and Ari never touches your laptop. It has no desktop and no mouse, just files and a command line, which turns out to be what most jobs need.

What that gets you:

  • Files of any type. Fill in a PDF form and sign it with a picture of your signature. Merge, split, rotate or shrink PDFs. Pull the text out of a scan. Make and edit Word, Excel and PowerPoint files. Convert, crop, trim and compress photos and videos. Ari saves results as new files, so your originals stay as they are.
  • Installing things. If a job needs a tool that is not there, Ari installs it on that machine. In testing I asked Ari to install a silly package, redeployed the computer service, and checked the package was still there afterwards. It was.
  • Running things. A quick command runs right in the chat. A download or a server can keep going between messages, and Ari checks on it later.
  • Building things. Ask for a small tracker or a dashboard and Ari builds it, with real records you can open and edit.

A job that takes a while shows up on a “Working on” card with the newest step as it happens and a Stop button. Stop a computer task and it pauses. Your files stay as they are, and Ari can pick it back up.

Everything lives in your Library, under Files. Open it, download it, upload to it, rename it. Deleted files sit in Trash for 30 days.

The browser

Ari drives a real browser. A browser that logs into sites, fills forms, clicks through flows, and grabs screenshots and downloads.

An invented example. You text: “book me the first open haircut slot on Thursday afternoon.”

  1. Ari opens the barber’s booking page.
  2. Ari finds the Thursday slots and picks the first afternoon one.
  3. Ari fills in your name and phone number.
  4. Ari stops at the confirm button and asks you.

You can watch all of it. The live view shows up right in the chat, and on a phone the page fills your screen when it is your turn to log in or take control. Two big buttons sit on it: Stop task and Take control. If you take control, you click and type on the same page Ari was using, then hand it back with one tap.

Some other things that fit the same shape: check whether an item is in stock, check an order status, file a form on a government site, renew something that expires.

It does not always work. Sites change, some pages fight automation, and a few things need a human. When Ari gets stuck, Ari tells you where, instead of pretending it worked.

Control

Giving an AI a browser and a computer only works if you stay in charge. Here is how I set that up.

Diagram

Approvals before anything that spends or sends. In the browser, before Ari buys, sends or submits something, you confirm it first. On the computer, Ari asks before using a saved key. The confirmation shows the exact amount and the recipient, and if either one changes, Ari asks again. I do not want to find out about a purchase after the fact. An approval is tied to one site and one session, and it expires. Nothing Ari can call is able to approve a request. Only you can.

Passwords and keys never go in the chat. Your website logins and your keys live on one Passwords page, in an encrypted vault. Ari’s browser fills a login only on the site it belongs to, and the value never passes through Ari. If a task needs a login Ari does not have, Ari sends a link to a page where you enter it. You never paste a password into a conversation. Ari’s computer can use a saved key with a program only after you say allow once, allow always, or deny, and you can revoke that later.

You step in for the human parts. Sign-in codes that arrive in your connected email get filled in for you. A code that comes to your phone, you text back or tap Fill, and Ari enters it without reading it. For a CAPTCHA, you decide per site whether Ari tries it, hands it to you, or asks each time.

Everything is visible. An activity list shows every step. You can stop any task at any time. When Ari finishes, the browser stays open and yours until you close it, so you can finish a checkout yourself if you want to.

The sandbox is walled off. Ari’s computer cannot reach any private network, mine or yours. It is a separate box with its own files and the public internet.

Why I built it

I kept running into the same gap. Reminders helped me remember the task. They did nothing to make the task get done. The form was still sitting there, and I was still the one who had to do it.

The whole point of Ari is that Ari manages your life, frees your time, and makes sure you reach your goal. A goal is rarely “remember to do the thing.” It is “have the thing done.” Doing it takes a computer, and I did not want to be the middle layer between Ari and one.

I also did not want to give an AI the keys to my own laptop. So Ari gets a separate computer, with a browser I can watch and a stop button I can press.

What is still rough

Ari is not finished. Browser tasks can stall, and a slow site makes a slow task. Some sites block automation outright. Ari can read a page wrong. That is why the approvals exist, and why the live view exists. Ari is artificial intelligence, not a person, and big decisions with money, health or the law stay with you.

Try it

Send Ari your first message at textari.chat on iMessage, Telegram, Discord, or the web chat. It is the same conversation everywhere. Then ask Ari for the thing you keep putting off.

The whole system is described on the Ari project page, and the reasoning behind the chat-first design is in why I built Ari. Ask Ari for anything. If it can be done on a computer or online, Ari can do it for you, and where you need to step in, Ari tells you.

✳ that's the note.

Webmentions

Letters to the editor

More from the notebook

← home saulutions.ca · field notes · written by Michael Saul