Get started

Response codes

Guard uses conventional HTTP status codes. The Gateway adds two intentional guardrail status codes (246 and 446) you can branch on.

Guard — HTTP status codes

Code	Meaning	When
200	OK	Request succeeded.
400	Bad Request	Malformed parameters (e.g. mixing last with from/to).
401	Unauthorized	Missing or invalid API key.
404	Not Found	Unknown resource or user.
422	Unprocessable Entity	Request body failed schema validation.
429	Too Many Requests	Rate limit exceeded.
500	Internal Server Error	Unexpected error evaluating detectors.
503	Service Unavailable	Data store temporarily unavailable (export routes).

Guard is fail-open by design

On the Guard API a flagged or deny result is returned with HTTP 200. Inspect the JSON body (flagged, deny) to make an enforcement decision — do not rely on HTTP status to detect a violation on /v1/guard. Error responses share a consistent shape: { "detail": "..." }.

Gateway — guardrail status codes

The Gateway signals guardrail outcomes through the HTTP status code itself, so you can enforce on the status without parsing the body. These are intentional, non-standard codes — branch on them directly.

Code	Meaning	Behaviour
200	Passed	All guardrail checks passed (or none configured). The provider response is returned as-is.
246	Flagged, not blocked	A guardrail check failed but the config did not deny, so the request proceeded. The provider response is returned with a `hook_results` object attached. Treated as success — use it to observe flagged-but-allowed traffic.
446	Blocked	A guardrail check failed and the config denied the request, so it was NOT sent to / returned from the provider. The body is `{ error: { type: "hooks_failed", ... }, hook_results: { ... } }`.

Both before-request and after-request guardrails can produce 246/446, and both work for streaming responses. Provider/LLM errors are passed through with the provider's own status code.

Branching on Gateway codes

Node.js

const res = await client.chat.completions.create(...).catch((e) => e);

// The OpenAI SDK throws on non-2xx; inspect the status to enforce.
if (res.status === 446) {
  // Guardrail blocked the request — read res.error.hook_results
} else if (res.status === 246) {
  // Flagged but allowed — response is usable; res.hook_results has detail
}