Admin guide · Platform

Guardrails

Policy rules evaluated on every request before it reaches the model. All guardrails assigned to a key must pass for the call to continue. Manage them under Guardrails.

Rules & activity

The activity strip counts triggered / blocked / flagged hits over the selected window; each rule has an Enabled toggle.

The Guardrails page with activity KPIs and a rule row showing type, action and enabled toggle. — Guardrails — rules with their action and enabled state.

Configuring a rule

A rule's behavior is defined by three things.

Field	Type	Description
Type	denylist · llm_judge	A `denylist` matches terms/patterns; an `llm_judge` asks a model to rule on the request. The type is fixed after creation.
Action on match	block · flag	`block` rejects the request; `flag` lets it through but records the hit.
Judge model & prompt	llm_judge only	The judge model and a judge prompt pulled straight from the Prompt Vault (optionally pinned to a version/tag), plus a timeout and max-input-chars.

The guardrail editor for an LLM judge: action on match, judge model and a judge prompt from the Prompt Vault. — Editing an LLM-judge rule — the judge prompt comes from the Prompt Vault.

tip · A rule says what to check, not where

The form has no surface to pick: a new guardrail is created assigned nowhere and runs only where something claims it. Assign it on an API key to evaluate that key's LLM calls, or on an MCP server (its page → Overview → Compliance) to evaluate that server's tool calls. The two assignments are independent and one rule can carry both — assigning it to keys never changes where it runs on the gateway, and the other way round.

warning · On MCP servers the rule runs twice

A guardrail assigned to an MCP server is evaluated on the tool arguments before they are forwarded and again on the result that comes back, for every caller that reaches the server — key or person, since gateway policy belongs to the upstream server rather than the credential. Blocking a result keeps it from the agent, but the tool has already run upstream.

EU AI Act enforcement

A separate, per-key policy that scores the prompt's EU AI Act risk level and blocks or flags high-risk or prohibited prompts.

It complements guardrails: guardrails encode your policy, while this compares live traffic against the risk level the key declared for itself. Enable it on the API-key form; the classifier behind it, and the reporting, live under EU AI Act.