Control who can access what

The golden rule

Assistants start with no access. Nothing — no skill, tool, or knowledge — is available until you allow it. You open things up deliberately.

There are two ways you control access, and they work together.

1. Grant skills, tools, and knowledge

When you share a skill, connect a tool, or add knowledge, you decide who gets it — a single person, a team, a department, or the whole org (the scopes again).

A grant is just "allow this for these people." Grant the Sales department your CRM tool; grant Engineering the code-review skill; keep a finance dataset to the finance team. Widen or revoke at any time, and changes take effect almost immediately.

2. Set guardrails with an access policy

An access policy draws a boundary around what an assistant may reach on the network — which external sites it can call, and which connected tools are off-limits. It's your safety net.

oc policies list
oc policies create --body '{
  "name": "default-egress",
  "domains": { "allow": ["*.example.com", "api.openai.com"], "defaultDeny": true },
  "mcpServers": { "deny": ["risky-tool"] }
}'
oc policies get <name>
oc policies delete <name>

defaultDeny: true means "block everything except what's listed" — the safest starting point.

Putting it together

Grants decide what an assistant can use. Policies decide where it's allowed to reach. Together they keep every assistant scoped to exactly what its owner needs — and nothing more.

Every change here is recorded in the audit log.