Run a website task every day with zero AI cost

Blog·By Rohit Kashyap·Published 18 October 2026

I build Dejo, and this post uses a feature that shipped in Dejo 0.1.2. Everything below is real output from my own machine.

A lot of browser work is the same small job every day: check a number on a dashboard, see if a status changed, grab this month's invoice. AI agents are good at figuring out how to do these the first time. They're a poor fit for doing them every morning, because every run is a new chat that reads the same pages again, and somebody has to start that chat.

Here's the setup I use instead: let the agent learn the task once, then let the operating system's scheduler run it, with no model at all.

Step 1: let your agent learn the task once

With Dejo added to your agent (setup takes a minute), ask for the job in your own words:

Go to github.com/microsoft/playwright and tell me how many issues are open.

The first time, the agent explores the page and saves the route with dejo_save. That's the only run that costs tokens. If the site needs a sign-in, do it now too: the agent opens a window, you sign in, and the session is kept. See signing in.

dejo sites shows what was saved:

github.com    count open issues on microsoft/playwright    2 steps  1 runs  health 100%

Step 2: run it from the terminal

dejo run replays the route with no agent and no model:

dejo run "count open issues on microsoft/playwright" --site github.com

  dejo  github.com  (warm run)
  task   count open issues on microsoft/playwright

  + step 1  open the microsoft/playwright issues page   [url]  1560ms
  + step 2  the exact open issue count for microsoft/playwright   [css:...]  501ms

  done in 2062ms  ·  2 steps from memory  ·  1 tool call  ·  0 model calls

  + the exact open issue count for microsoft/playwright:  Open 175 (175)

Two seconds, zero model calls, and the answer on the last line. You don't have to type the task exactly as it was saved; close wording is matched.

Step 3: get output a script can read

For a scheduler you want something a script can parse. --json prints one line:

dejo run "count open issues on microsoft/playwright" --site github.com --json

{"outcome": "done", "ok": true, "site": "github.com", "task": "count open issues on microsoft/playwright", "answers": {"the exact open issue count for microsoft/playwright": "Open\n175\n (175)"}, "saved_files": [], "duration_ms": 1050, "model_calls": 0, "reason": "", "repair": null}

The exit code tells you what happened without reading the JSON at all:

Step 4: schedule it

On Linux or macOS, cron. Cron runs with a very short PATH, so use the full path to dejo (find it with which dejo):

# every day at 9:00, append the result to a log
0 9 * * * /home/you/.local/bin/dejo run "count open issues on microsoft/playwright" --site github.com --json >> /home/you/dejo-daily.log 2>&1

On Windows, put the command in a small script and let Task Scheduler run it:

# C:\scripts\daily-check.ps1
dejo run "count open issues on microsoft/playwright" --site github.com --json |
  Add-Content C:\scripts\dejo-daily.log

schtasks /Create /SC DAILY /ST 09:00 /TN "Dejo daily check" /TR "powershell -File C:\scripts\daily-check.ps1"

The run is headless, so no window opens and it doesn't need a screen. Signing in is the only part that needs one, and you did that in step 1.

Step 5: only hear about it when something's wrong

A log is fine for numbers you check now and then. For a job you care about, act on the exit code. A small bash example that writes the answer on success and a warning otherwise:

#!/usr/bin/env bash
out=$(dejo run "count open issues on microsoft/playwright" --site github.com --json)
code=$?
if [ "$code" -eq 0 ]; then
  echo "$(date +%F) $(echo "$out" | jq -r '.answers[]' | tr '\n' ' ')" >> ~/issues.log
else
  echo "$(date +%F) dejo stopped with code $code: $(echo "$out" | jq -r '.reason')" >> ~/dejo-problems.log
fi

Swap the last echo for whatever reaches you: an email, a Slack webhook, a desktop notification.

Things to know before you rely on it

• The browser profile is shared with your agent. Chrome allows one running browser per profile, so if a scheduled run starts while your agent has a sign-in window open, one of them can't open the profile. Schedule it for a quiet time, or give the scheduled job its own profile with dejo --profile scheduler ... and sign that profile in once.
• Passwords never go into the saved route. If a step fills a password form, put the password in an environment variable (DEJO_SECRET_<SITE>_<FIELD>) that the scheduled job can see, or in ~/.dejo/secrets.json.
• Sessions still expire on the site's schedule. When one does, the run exits with 3 and you sign in again. Nothing else changes.
• If the site changes, the run exits with 1 and names the step. Ask your agent to run the task once; it repairs that step, and the scheduled job goes back to costing nothing.

The point of all this is that the expensive part, an AI figuring out a website, happens once. Everything after that is a two-second script. If you want the numbers on what that saves, they're in why your agent re-reads every website.

FAQ