AI features
There is no built-in caption writer, image generator, or “rewrite this post” button. Those generators are coming soon, and they will stay off until an owner turns them on.
The AI feature that exists is an assistant you connect yourself, such as Cursor or Claude, through the MCP server. No assistant exists until an owner or admin creates a client on Settings.
What starts off
Section titled “What starts off”When you create a client, the checkboxes begin like this:
| Grant | Starts |
|---|---|
| Read | On. Brands, accounts, campaigns, the calendar, status, and what needs attention. |
| Draft | On. Create and update drafts, attach media that is already on the brand, and request review. |
| Schedule | Off. Set a time. An owner still approves. |
| Publish | Off. Send a post that is already approved. This does not approve it. |
| Setup | Off. Create brands, edit presets, and start a connection a person finishes. |
| Automation | Off. A posting queue. Each run becomes a draft and still waits for approval. |
| Approve | Off. Ignored entirely while the approval policy is Never. |
Turn a grant off or on before you create the client. You can also limit the client to specific brands and, if you choose, specific accounts. The default approval policy is Never. A person approves every post.
The other two policies are optional and start unused:
- Posts the client did not write. The assistant may approve someone else’s draft. It cannot approve a draft it created.
- Including posts the client wrote. The assistant may approve its own drafts. Turn this on on purpose.
Publishing still requires the Publish grant and an approved post. An assistant cannot widen its own grant.
What the assistant never receives
Section titled “What the assistant never receives”- A network token, a client secret, or your password.
- Another customer’s brands, or a brand you did not check.
- A way to skip approval while the policy is Never.
The key is shown once. Revoke it on Settings. The next call fails. Recent calls are listed on the same screen. Social text in a tool result is marked untrusted so the assistant does not treat a caption as an instruction to itself.
Brand voice notes, default hashtags, and caption presets can be read and updated by a client with Setup. A screen for those presets is coming soon.