Prompts: Change your prompts. Without redeploying your app.

Saving a new version also sets it as the active version. If a version is pinned to production, that pinned version overrides the active version at request time—so activating another version alone will not change production until you move or clear the pin. Owners and Organization Admins can pin, unpin, or activate versions in the dashboard; once your application is integrated, switching or rolling back requires no application redeploy.

You can pass nrouter_prompt_template_id and nrouter_prompt_variables in extra_body on the request, or bind a template in the dashboard by virtual key, team, or organization. When no template ID is sent on the call, assignments resolve in order of specificity: key, then team, then organization. The gateway strips the extra_body selector fields before forwarding; on chat completions, it prepends the rendered prompt as a system message ahead of your existing messages array (and sets or prepends system or instructions on other text routes).

Templates substitute double-brace placeholders ({{name}} and {{ name }}) with string, number, or boolean values—without loops, conditionals, or template filters. Request variables override default assignment variables, and authenticated system variables (org_name, model, timestamp, and user_id) are applied last so caller input cannot overwrite them. Rendering runs before credit reservation and provider dispatch: if any placeholder is left unresolved, a variable is not a string, number, or boolean, or the rendered prompt exceeds the 256 KiB ceiling, the gateway returns HTTP 400 and charges $0.00.

An A/B test splits non-streaming traffic across two or more prompt or model variants whose weights sum to 1.0 (±0.01) and records latency, token usage, and cost for completed calls. Assignment is deterministic per virtual key (or user ID), so a single shared API key stays on one variant rather than splitting its individual calls 60/40. Streaming requests selected into a running experiment are refused with HTTP 400 before provider dispatch; requests that are not selected into the experiment can still stream normally.

Prompt management is controlled by an organization-level setting. If prompt management is disabled for your organization, regular inference calls proceed normally, while requests that pass nrouter_prompt_template_id or nrouter_prompt_variables are refused with HTTP 400. Once enabled and connected via extra_body or a scope assignment, your team can iterate on templates without redeploying application code.

Templates, versions, and assignments are scoped to your organization, and gateway prompt execution records store metadata only (such as template ID, version number, model, and variant name) without saving prompt or completion text. When a request runs, the rendered prompt is sent to the selected upstream model provider to generate the response, and any optional request-content logging follows your organization data policy.