What you can do with Dejo
The short answer: use Dejo for website tasks your AI agent does again and again. The agent learns the task once on the real site; every run after that is one tool call with zero model calls, and if the site changes, only the step that moved is learned again.
Each example below shows the steps your agent would send the first time. You do not write them yourself: you ask your agent in plain words, and it sends steps like these. They are here so you can see what Dejo saves.
A number you check every week
Last week's sign-ups, a balance, open tickets, a status. Your agent opens the page and reads the value once; after that, asking again is one call.
go https://dashboard.example/signups
read css "#weekly-signups" as signups
To run it on a schedule with no agent at all, dejo run replays the saved task from cron or Windows Task Scheduler. See run a website task every day with zero AI cost.
Tests that check every step
An end-to-end flow (sign up, search, add to cart) learned once and replayed as a test. Every saved step carries a check that it landed, so a replay is a test by itself: it stops at the first step that did not land and says which one. A changed button is found another way before anything fails, and a cookie banner or toast in the way is cleared.
go https://shop.example/products
fill label "Search" "mug"
press Enter
click link "Blue mug" first
expect text "Add to cart"
How it behaves when the site is different on the next run: when the site changes.
Your site on an iPhone, after every deploy
Open your own pages as an iPhone 13, check the layout and take a screenshot, then repeat the same check after each deploy with no model calls.
emulate device "iPhone 13"
go https://your-site.example/pricing
screenshot full
read css "nav" as menu
The full walkthrough, with real runs on five pages: test your website as an iPhone from Claude Code.
Admin, customer and vendor accounts side by side
A marketplace, a clinic, a school system: one site, several roles. Each role gets its own signed-in profile, so the agent can check what a customer sees and what the admin sees in the same session, without signing in and out.
dejo_profile(name="vendor")
dejo_run(site="market.example", task="list my open orders")
You sign in to each profile once, yourself, in a real window. Dejo also notes which account each task was learned as, and stops if a different one is signed in. See signing in and teams with more than one role.
Monthly admin work behind a sign-in
Download this month's invoice, export a report, update a record. A step that sends, pays, deletes or submits is marked with !, so it never runs by accident and the reply always names it.
go https://billing.example/invoices
click link "November 2026"
click! button "Download PDF"
Prices, stock and status on sites with no API
Read a price or "in stock" from a supplier, a status page or a listing, every day, without paying a model to read the page each time.
go https://supplier.example/item/4471
read css ".price" as price
read text "In stock" as stock
Many fields at once come back as JSON with read fields, and a whole table with read table. See the tools reference.
Your own Chrome, or a cloud browser
Point Dejo at a browser that is already running: your own Chrome with your sign-ins, or a cloud browser. Set one setting in the MCP server's environment:
DEJO_CDP_URL=http://127.0.0.1:9222
Dejo works in a tab of its own, never one you have open, and leaves the browser running when it is done.
A video of a flow, for a demo or a bug report
Record what happens as a video and a trace that opens in Playwright's trace viewer.
record video trace
go https://your-app.example/onboarding
click button "Next"
record stop
FAQ
What is Dejo good for?
Website tasks your AI agent does more than once: a weekly report, a dashboard number, a test flow, a check after every deploy, admin work behind a sign-in. The first run is learned; every run after is one tool call with zero model calls.
When is Dejo not worth it?
A one-off lookup on a single page. The savings come from repetition, so a task you will never run again costs about the same as with any other browser tool.
Which AI agents can use these?
Any agent that can run an MCP server: Claude Code, Codex, Cursor and others. The same saved task also runs from a terminal or a scheduler with `dejo run`, with no agent at all.