Your agents. Your rules.

Approval before action

Every job carries a setting: watch only, ask me first, or run on its own. Approval policies define which actions require a decision before execution.

Money and deletion never graduate automatically. Those always ask.

A spend cap that actually stops

Your limit is checked before each paid action, not totalled afterwards. Hit it, and agents pause and tell you. That applies to everything that costs money (text, images, search, browsing, connectors), not just one category.

A runaway bill isn’t something we’d rather avoid; it’s something the system won’t permit.

Your data is your workspace

Each customer’s data lives in its own isolated workspace, enforced at the database level rather than by application code that could be bypassed. Access is scoped by who you are: owners see their company, team members see their own work.

Your data is exportable whenever you want it. Nothing we produce for you is locked in.

Least privilege for every agent

An agent gets access to specific tools and specific actions, not to your whole account. A finance agent might read Stripe but need approval to refund and be unable to send email at all. Permissions are per-agent, not per-account, and they’re enforced outside the AI, so the model does not decide its own authorization.