zwrm commands — deploy, read logs, create databases. Capabilities decide which of those the token is allowed to do.
Default
An agent gets every capability exceptdestroy. It can build, ship, and manage; it cannot irreversibly delete an app, machine, database, or volume.
Available capabilities
Scopes areresource:action pairs, plus a few atomic verbs:
Transferring an app to another organization is not grantable at all: no agent or scoped token can move an app out of your org.
What the agent sees
A call outside an agent’s capabilities fails with403 token is not authorized for this endpoint. The platform instructions tell agents that a 403 here is a policy limit rather than a bug — they should report it and ask you to grant the capability, not retry or work around it.
Agent tokens are also pinned to their agent and organization: an agent’s token only acts as that agent, inside its org.
The pin confines agent-addressed routes. Org-level resources — the skill library, schedules, triggers — resolve their target from the request rather than the URL, so an agent granted
skills:write or schedules:write can manage those across its own organization, not just its own agent.Scoped API tokens
The same vocabulary works for your own tokens, so a CI job or dashboard gets exactly the reach it needs:
A scoped token is confined to its organization — it cannot reach another org even if you’re a member. Unscoped tokens keep behaving as before, carrying your full access.
Audit log entries record which credential made each call, so a scoped token’s activity is attributable to that token rather than just to you.