DocsTeams
Governance policy explained for admins
Everything a team's agents can do runs through one policy that admins set in Settings > Governance. It applies everywhere at once - to the assistant, to agents, to automations, and to anyone calling the gateway - so there's exactly one place to look for "what is this team allowed to do."
What the policy covers
- Models & providers. Which AI providers and models are allowed, and which one is the default for a new agent.
- MCP & skills. Which external MCP servers and skills members may connect, plus any skills you pin so every member's assistant has them automatically.
- Tools. Allow and deny rules for what agents can touch, down to specific arguments (for example, only allowing file writes inside a particular folder), and which actions require an approval before they run.
- Budgets. A monthly spending cap for the whole org and, optionally, a smaller cap per person.
- Rate limits. Caps on how much any one member or the org as a whole can call models in a given window.
- Guardrails. Automatic redaction of personal information and secrets, and screening for prompt injection attempts, applied before anything reaches a model.
- Routing. Fallback behavior when a preferred model is unavailable or over budget.
- Caching. Whether repeated or similar requests can be served from a cache instead of calling a model again.
- Approvals & the two-person rule. Which actions need a second admin's sign-off, and how the team's "break glass" emergency override works.
- Data residency. Restricting which regions a request may be routed through, where a provider supports it.
- Speculation. The ceiling on speculative (branch-prediction) execution described in Speculation.
Trying a change before you make it
Before you save a policy change, you can preview what it would have blocked over the last 30 days of real activity - so you find out about a policy that's too strict (or too loose) before it affects anyone. Every past version of the policy is kept, with a diff and the ability to restore an earlier one.
Per-member overrides
Most policy applies org-wide, but an admin can override specific settings for one member, and admins themselves can be exempted from restrictions that shouldn't apply to the people managing the org.