Skip to content

Your first chat

Your first result takes about three minutes. When you sign in for the first time, Bartimaeus offers a Get started card with three short steps — you finish by watching your agent actually do something. The card is skippable and never blocks anything; if you skip it, it stops asking.

The three steps

1. Connect your model key. If you have not added a key yet, step one opens the key wizard and returns you here when it is done. Already keyed? This step ticks itself. (Bring your AI key covers the wizard in full.)

2. Pick a first task. Three starters are offered, all designed to work without connecting any apps — so your very first run cannot dead-end on a missing connection:

  • Draft an email
  • Plan my day
  • Turn a recipe into a shopping list

3. See it work. Tap Run. Your prompt goes to your model provider with your own key; the reply comes back through Bartimaeus. Watch what happens in the conversation.

The result moment

When your first run settles cleanly, the conversation shows a one-time card called What just happened:

Your prompt went to the model provider with your own key. The reply came back through Bartimaeus — nothing was sent or changed anywhere else.

It also offers the natural next move: Make it a routine. One tap opens a routine with your task's name, description and prompt already filled in, so "ask once" can become "happens every Tuesday" without any setup work. (Routines are a Premium feature; Routines explains more. Done dismisses the card.)

Asking for things

After the first run, it is just conversation. Some shapes that work well:

  • Be concrete: "Draft a polite email to the plumber about Thursday's visit."
  • Give it context you would give a person: dates, names, the outcome you want.
  • Ask for a plan first if you are unsure: "Tell me how you would approach this before doing anything."

The moment you connect apps — Gmail, Calendar and the rest — the same conversation can reach into them. That story lives in Connecting apps.

While your agent works

  • A working indicator shows a run is in progress. If a request takes more than two silent minutes, it is replaced with a plain message that something on our side is struggling — not left spinning.
  • The send button becomes a stop button while a run is going. Tapping stop ends the run immediately; you can re-prompt straight away. If the agent was parked on an approval card, that card is cancelled with the run — nothing happens after a stop.
  • If the agent needs your permission for a sensitive step, the run pauses, you get a notification, and an approval card appears in the conversation. Nothing goes ahead until you decide. See Approvals and the safety model.
  • Everything leaves a receipt. What was asked, what was done, what changed — visible in the conversation where it happened.

What not to expect

  • It is not always on. Your agent works only while there is work: your prompt, a scheduled run, or a trigger. Sessions stop when the work is done.
  • Results vary by model. Your agent is only as sharp as the model you connected and the clarity of the request — both are under your control.
  • The Free plan has a time budget. 30 minutes of active compute a month. It is enough to see what an agent does with your own accounts; the plans page compares it with Premium.

Where next