Sidekick Preview
Say what you want done. Approve what it proposes.
The shell you talk to. It finds what needs attention across your apps and proposes actions; it does not own the model, you do.
How a task goes
- You describe the task in plain words, for example "chase the three invoices that are overdue and tell the account owners".
- Sidekick finds what is involved across your apps and proposes the actions, each with the evidence it used.receipt: the proposal and its sources
- You approve exactly what will happen. The approval is bound to the content you saw; if it changes, you are asked again.receipt: the approval
- It does the work and checks the result against the system that holds the record, and tells you what actually happened.receipt: the verified result
What it does not do
- It does not act without an approval on anything that sends, pays or publishes.
- It does not own the model. You choose it and pay for it: a local model, or a hosted provider you already use. Your agent, your key.
- It does not replace your apps. They stay where they are and keep their own records.
Where it stands
Preview Sidekick works in the default install, on sample data until you connect your own accounts. The task-to-approval-to-verification flow above is what it is built to do; how well it understands an open-ended request depends on the model you give it, and small local models can draft but should not be left to approve anything.
What is shipped and what is not is listed on what is available today and trust.
Try it live. Projects and Sidekick as a read-only public demo. Workflows can be listed and their steps explored; composing, validating and launching are disabled here.