This walks through a first real task: attach a file, let the agent run code over it, and export what it produces.
Add a model credential
Open Settings → Providers, turn on the provider you want, enter its credential, and press Save configuration. Providers take an API key; the OpenAI-compatible slot also takes a base URL.
If you have a ChatGPT Plus or Pro subscription, the OpenAI section offers Sign in with ChatGPT instead of a key.
Pick a default model
Go to Settings → Models. Under Chat, choose the provider you just configured and a model it serves. Leave Background work automatic unless you have a preference — it resolves against whatever you have credentialed.
Start a chat and attach a file
Press New chat. In the composer, open the Tools menu (the + button)
and choose Attach files, or type @ in the message box to search files
and folders inline.
Attached files show as chips above the message field.
Ask for something that requires work
Not "summarize this" — something the agent has to compute:
Read the attached CSV. Chart monthly revenue by region and
save it as a spreadsheet with the chart embedded.
The agent will write and run a script. The first time it runs a command, you get an approval card asking whether to allow it. Choose Yes, run it once, or one of the "always allow" options to stop being asked for that kind of command in this chat.
Take the file
Files the agent writes into its output/ directory are published as outputs.
They appear in a strip above the transcript and in the Outputs panel,
reachable from the activity chip in the chat header.
Open one, then use Save as… to write it wherever you want with a native save dialog.