Skip to main content
Guardrails are organization-level rules — a name plus an optional description — that every agent in the organization must follow on every response and tool call. They are injected into each agent’s system prompt at call start. Guardrails apply to all agents in the organization; there is no per-agent assignment and no feature gate.

​
Limits

  • Up to 5 guardrails per organization. Creating a sixth returns 409 limit_reached.
  • name: 1–100 characters, unique within the organization.
  • description: up to 1000 characters.

​
Workflow

  1. Create guardrails for an organization (e.g. “No PII — never reveal or confirm a caller’s full account number, SSN, or payment card details”).
  2. Toggle enabled to pause a rule without deleting it — disabled guardrails stay listed but are not injected into the agent prompt.
  3. Update or delete rules as policies change; changes take effect on the next call.

​
Endpoints

The API exposes exactly the same guardrail fields you manage in the dashboard — nothing more. There are no internal or secret fields on a guardrail.

​
Guardrail object

GET and POST return the full guardrail object (inside a data wrapper); PATCH and DELETE return { "success": true }.

​
Errors

Errors use one of two JSON shapes with the matching HTTP status:
  • Request-schema failures return { "error": "Invalid request data", "details": [...] } (400), where details is the list of failed field validations.
  • All other errors return { "error": "<code>", "error_description": "<message>" }.