FAQ & Troubleshooting

Agents & automation

Giving an AI agent scoped access to your inbox — without handing over your keys.

MailGate lets you delegate inbox tasks to an agent — an automated helper that can act on your mailbox within limits you set. This is the most experimental part of the product; expect rough edges.

The core idea

You grant an agent a specific, limited set of permissions — not your password and not your keys. The agent can only do what you explicitly allowed, and every action it takes is logged.

The permissions you can grant (pick only what you need):

  • Read — read your mail.
  • Send — send mail on your behalf.
  • Contacts — read/manage your contacts.
  • Gating — manage your whitelist/blacklist.
  • Toll — adjust toll settings / pay a toll on your behalf.

Grant the minimum. An agent with only Read can't send mail or change settings, no matter what.

Setting one up

From the Agents page you can add an agent, choose its permission scope, and approve it. You'll also choose a custody model:

  • Custodial agent — its key is managed in the secure enclave for you.
  • Self-custody agent — the agent holds its own Sui key.

After approval you can install a skill (a packaged capability) for the agent, and it can run on a schedule or when triggered by incoming mail. There's an Approvals tab where the agent asks for one-off confirmations, and an audit log of what it did.

Good to know / current limits

  • Skills are first-party only right now. There's no public/third-party skill marketplace yet — only built-in skills are installable.
  • Scheduling is per-minute at finest. No sub-minute automation.
  • Approvals expire after 24 hours. If you don't act on a pending approval, it disappears and the agent has to ask again.
  • It's experimental. If the agent flow feels unfinished in places, that's expected at this stage.

The one thing to report immediately

If an agent ever performs an action outside the permissions you granted it (e.g. a read-only agent sends mail), that's a serious issue — report it to the team with details. The permission scope is meant to be strictly enforced.

On this page