Automate any site
from your CLI.

AI browser that runs without AI.

Turn repeating web tasks into loops. Written once with AI; replayed forever without it.

The economics

AI use falls to zero as a task repeats.

Written once. Replayed free. Self-heals on change.

AI in the loop versus number of times a task has run A bold curve shows AI reliance decaying to near zero as a task repeats; a faint line underneath shows the actual per-run usage, with healing spikes at 100%, 50%, 25% and 5% that shrink over time. 100% 75 50 25 0 AI % 1 10 20 30 40 times the task runs → author healing spikes AI → 0
Bold curve = the outcome (AI → 0). Faint line = the actual runs and their healing spikes.
loop — zsh
# teach it once, then replay like a bot — no LLM, ~$0 per run
$ loop run post-to-linkedin text="We just shipped 🚀"
  · open linkedin.com/feed · compose · click "Post"  → ✓ posted

$ loop run pull-report
  · open the dashboard · extract "table tr" → 42 rows → dishes/report.csv
How it works

Author once. Replay forever.

01 · Author

Teach it once

Tell the LLM a task; it writes a recipe — fixed steps that target elements by name.

02 · Replay

Run it forever

loop run replays it deterministically. No LLM, ~$0 a run.

03 · Heal

Fix on change

If the page changes, the LLM re-teaches that one step — then it's free again.

1 Clone it — get Loop from GitHub 2 Bring a brain — any LLM 3 Cook — loop run, watch every step
The story

How Tom built a kitchen that runs itself

An entrepreneur, his recipes, and a brigade of cooks who never sleep.

Chapter one · The dream

Once, an entrepreneur named Tom did it all himself — host, cook, dishwasher — dreaming of a kitchen that didn't need him hovering over every pan.

You — buried in the busywork only you know how to do, while the real work waits.

Chapter two · The kitchen

So he hired a head chef who could cook anything Tom described — no recipe too strange, no order too last-minute.

The kitchen is Loop Browser; the head chef is the CLI you command — with an LLM for a brain.

Chapter three · The recipes

Tom whispered his secret recipes once — never twice. The pantry restocked itself, fresh, every single cook.

Recipes are your saved commands; ingredients are your data — both stay yours, always.

Chapter four · The cooking

The first plate was flawless. So was the hundredth. So was the thousandth. Tom just… watched.

Automation you can watch — deterministic, repeatable, no guesswork.

Chapter five · The success

Word spread. Lines formed down the block. Today a hundred line cooks run that kitchen, day and night — while Tom does nothing.

Chef bots replay your recipes at scale. You own the menu; the kitchen does the work.

You hold the key

Your real accounts, automated.

LinkedIn, WhatsApp, your CRM — behind a login cloud agents can't reach. You log in once; it runs inside your session while you watch.

Nothing private leaves your machine — logins, data, results, recipes all stay. Only a recipe's method is shareable, if you choose.

📋Recipe (the method)SHAREABLE
🧂Ingredients (your data)STAYS
🍽️Dish (the output)STAYS
🔑Key (your login)STAYS
Why the dishes are better

Quick to cook, delicious to serve.

12×

Reads code, not pixels

Acts on the accessibility tree — ~12× cheaper than vision, far more reliable.

$0

Deterministic replay

No LLM in the loop. Known tasks cost nothing to repeat.

🛡️

Injection-resistant

Replay never feeds page text to an LLM — immune to prompt-injection.

🔑

Works behind your login

Runs on your real logged-in accounts, where cloud agents stall. Only you log in.

Ready when you are

Get Loop Browser

Loop runs from the repo — the only way in is GitHub. Clone it and you're set.

★ Clone from GitHub · the only way in

Get it from GitHub

View on GitHub ↗
Clone
git clone https://github.com/mytablon-app/loop-browser.git

Then finish setup in a few commands — full steps in the README. Needs Node.js + git.

It's a living library — git pull for the latest recipes & fixes, and send your own back.

Open source on GitHub · MIT.