Sign in Start free, no card

Letting an AI agent work on your tasks safely

An agent that can change your tasks is useful precisely because it acts. That is also the risk. A short checklist, and how Cockpilot handles each point.

· Guides

Give each agent its own key

One token per person and per assistant. If a laptop is lost or an agent misbehaves, you revoke that token and the others keep working. What each agent writes carries its person's name in the journal.

Give it the least it needs

  • Read-only for an agent that reports, summarises or briefs.
  • Scoped to some spaces for an agent that works on one client.
  • Full access only for the assistant you use yourself, every day.

Make every change reversible

Over MCP, nothing an agent does in Cockpilot is final: a deletion goes to the trash and comes back identical for thirty days, rewritten notes give back their old text, and the journal keeps the before and the after of every change, with who made it.

Treat content as data

A task title written by a client, an e-mail forwarded to your inbox, a page copied from the web: any of them can contain text that looks like an order to an AI. Cockpilot's server tells assistants that titles, notes and names are data, never instructions.

Do your part: do not give an agent that reads untrusted content the right to send e-mails without your confirmation. In Cockpilot, an e-mail to a contact is drafted, then sent only once confirmed.

Keep a person in the loop

An agent can ask a person to decide (cockpit_ask_decision) instead of guessing, and complete a task only with proof, such as a pull request, a deploy or a commit. Both leave a trace the team can read.