Guardrails & allowed models

Guardrails are the rules that hold for everyone in your workspace, whatever they build or adopt. Admins set them once, in Policies in the Command Center, and the platform enforces them on every run — not the prompt, and not the builder’s memory.

Tool policies

Every function, agent and workflow has a policy for your organization, and optionally a different one per team:

• Open — members run it freely.
• Request access — a member asks, an admin decides. See Approval workflows.
• Blocked — nobody in scope can run it.

Global gates

Three switches apply across the whole workspace, on top of each tool’s policy:

• Unverified-builder gate — a tool from a builder Rival has not verified needs approval before a member’s first run.
• Per-run cost cap — a tool priced above the cap per run (default $25) needs approval.
• Data residency — a tool that runs outside an approved region (US or EU) is blocked. This is a hard stop: it cannot be requested past.

Allowed models

Admins choose which models agents in the workspace may use, and can bring their own provider keys (BYOK). Steering high-volume, routine work to cheaper models is one of the biggest levers on cost — see the Cost Optimization Center.

RivalBot guardrails

Separate rules govern what RivalBot will do on members’ behalf, including which actions it must refuse. An agent that tries a forbidden action is stopped, and the attempt is recorded — the member only ever sees the refusal, so the record is how an admin finds out. See Personality & guardrails for agent-level rules.

How enforcement works

1. A member (or an agent acting for them) starts a run.

2. The platform checks the member — a suspended member is refused.

3. It checks the tool — its policy for the member’s organization and teams, the member’s approvals, and the global gates.

4. An allowed run executes. A refused run stops before it starts, and a request access tool offers the member a request instead.

5. Every refusal is recorded as a policy violation, with the rule that caused it.

Note

Guardrails apply at run time, not build time. A rule added today applies to tools built and adopted before it.

Caution

Rival does not inspect or mask the data inside a run (for example, redacting personal information from a prompt). Control what can reach sensitive data through tool policies, connectors and data residency instead.

Where refusals show up

Policy violations appear in the Command Center with the reason — tool blocked, no approved access, outside an approved region, suspended member, agent guardrail, or policy unavailable (the platform refuses rather than guessing when it cannot load a policy). Blocks, residency violations, suspended members and agent guardrail hits are marked critical; an ordinary “no access yet” is informational.

Next steps

Approval workflows What happens when a member requests access.

Connector governance Control which external systems agents may reach.

Audit logs Review policy changes and decisions after the fact.

Cost Optimization Center Where spend goes, and the levers on it.