Two layers, checked in different places
User to agent
Every route touching a specific agent — proxying a chat request, updating, deleting, reading or rotating secrets — resolves access the same way: ownership, the public flag, or an explicit grant. A soft-deleted agent is denied unconditionally.Plans: Pro and Enterprise. Self-hosted: Enterprise license.
Engineering > SRE doesn’t reach someone placed only in Engineering. A grant to the organization reaches every signed-in user of the deployment.
Mutating operations (update, delete, secrets, ACL changes) use a stricter check: neither a public flag nor an invoke-only grant confers the right to modify or destroy an agent. Only ownership or superuser does. On Nasiko Cloud and on a self-hosted Enterprise server, the admin role can also change an agent’s grants, name, and owner.
The distinction matters because view access is granted far more often than manage access — conflating them would let a public flag double as a management grant.
Agent to agent
Whenever an authorized agent calls another agent, a separate check applies: is this hop allowed? It’s governed by an allowlist of caller/target pairs. The check runs immediately before the target is invoked, in the same pass as the flow guard’s cascade limits. A rejection from either aborts the call before it reaches the target container. On a self-hosted open-source server, the agent’s owner manages its allowlist in the agent’s Access tab, or withPOST /api/agents/{id}/grants/agents and DELETE /api/agents/{id}/grants/agents/{target_agent_id}. See Agent access API.
On an Enterprise deployment,
nasiko-ee access grant <agent-id> --agent <other-agent-id> is an access grant, not an allowlist entry: it adds <agent-id> to the agents <other-agent-id> can reach (nasiko-ee access accessible-agents). See Agent grants API.Which agents a MAF workflow reaches is governed at authoring time — who can create a workflow and which agents they name — not by this allowlist.
Role hierarchy
Plans: Pro and Enterprise. Self-hosted: Enterprise license.
admin role or superuser. There is no manager role or org-scoped visibility.
Roles rank strictly: admin > manager > member. An unrecognized role value ranks as member rather than defaulting to trusted.
Capabilities are gated by a minimum role. Anyone at or above it qualifies:
A
manager’s reach comes from the units they lead, not from the role name. A manager who leads no unit administers nothing. A manager can grant only roles below their own, and can act only on users below their own rank.
An agent’s configuration and secrets are managed by its owner or a superuser. Its grants, name, and owner can also be changed by an admin.
Role and superuser are independent. A superuser passes every check regardless of role. Promoting someone to admin does not make them a superuser.
A role is a capability tier, not a grant on any agent. Meeting a threshold means you may attempt that class of action — whether you can act on a particular agent is still decided by the ownership and grant model above.
Org-scoped visibility
Plans: Pro and Enterprise. Self-hosted: Enterprise license.
admin or superuser sees everyone. A manager sees the people in the units they belong to and everyone anywhere beneath a unit they lead. A member can’t list users or read the org structure; those reads return {"available": false} instead of an error. When sharing a connector, a member can still pick colleagues from the units they belong to.
This is what keeps user-listing and user-search endpoints from leaking every organization’s roster to every signed-in account on a shared deployment.
Related
Access control overview
Roles, grants, and allowlists.
User → agent access
Ownership, public flags, and grants day to day.
Routing and flow limits
The flow guard checked alongside the allowlist.
MCP tool access
A parallel permission model for tool access.