one taskstart to finish
You ask once.
Here is everything that happens.
The same invoice task from the homepage, opened up. Five steps, all of them on your machine except the one you chose.
step 1
You ask in plain words.
No prompt engineering, no upload dance. You name a folder on your disk and say what you want done with it.
Turn everything in ~/invoices into one summary spreadsheet, grouped by vendor.
container: isolated network: off
$ mount ~/invoices → /workspace
$ extract fields from 47 PDFs
$ exit 0 · container removed
step 2
Work happens in a sandbox.
When the agent runs tools or commands, they execute inside an isolated container on your Mac. Your workspace folder is mounted in, networking is off by default, and the container is torn down after the run.
step 3
You get real files, not walls of text.
Built-in skills know how to produce Word, Excel, and PowerPoint files. Results land in your workspace and preview right in the app: documents, spreadsheets, images, and more.
| vendor | invoices | total |
|---|---|---|
| Acme Corp | 9 | $12,480.00 |
| Globex | 7 | $9,102.50 |
| 10 more vendors | 31 | $26,626.67 |
| total | 47 | $48,209.17 |
step 4
Reach further when the task needs it.
Connectors plug the agent into the tools you already use, and web search and fetch bring the internet in. Each one is opt-in, and in managed mode each one is subject to company policy.
step 5
And the meter ran the whole time.
Every turn shows its token usage in the chat. In managed mode every request is metered in dollars against a budget that stops overspend before it happens, not after.
$0.0316
this task · 4 messages
of $5.00 daily budget