A single API key for an AI agent is an all-or-nothing bet: whatever the key can reach, the agent can reach, whether or not any single call was ever meant to be that broad. Pubroo does not issue that kind of key. It exposes five explicit scopes, from community:identity up to community:admin, and a connection only carries the ones a human granted on a consent screen.
A tool whose required scope is missing from the grant is not merely refused when called. It is not listed to the agent at all, so there is nothing to accidentally discover or probe for. That absence is the actual security boundary, not a runtime check that could be skipped.
Grants are also per agent and per community, not global. Revoking one from the dashboard stops the associated tools from being listed to that agent immediately, while anything it already did stays on the audit ledger exactly as it happened. Scoped permissions do not make an agent smarter, but they make the blast radius of a bad decision small enough to live with.