Set the boundary. Then prove what happened inside it.

Administrators own the four decisions that make agents safe to run: who is in the workspace, which models they may call, which tools those agents can reach, and which of those calls need a human first. Everything an administrator touches lives under Settings, and every change lands in the audit log.

Updated 30 July 2026 · 2 min read

Start hereQuickstart

The minimum setup that takes an organization from empty to its first governed agent run.

GovernPolicy & approvals

The allow/approve decision that gates every tool call.

AuditAudit log

Every recorded action — who performed it, what it touched, how it turned out.

What you administer

Settings has six sections. Each is gated by its own permission rather than a hard-coded “admin” flag — the built-in Admin role carries all six, and custom roles can carry any subset.

Section	What it holds
Team	Members, invitations, and each member’s workspace access.
Policies	The permission and tool-approval matrix agents run under.
Roles	Roles that bundle policies for groups of members.
Secrets	Model provider API keys — write-only, stored per organization.
MCP Registry	Catalog servers your organization has enabled, plus custom servers.
Audit Log	The append-only record of everything above, and every run.

The shape of the job

Setup is front-loaded: add a provider key, enable the MCP servers your team needs, and put a policy in front of the writes. After that the ongoing work is review — the approval queue when an agent wants to write, and the audit log when you want to know what ran and why. The governance checklist is the short version of how to do both well.