Configure an AI proxy

An AI proxy provides a gateway between Aviator and AI providers. The proxy is a separate bundle from the PPM application and runs on customer-managed infrastructure.

Overview

Administrators use the AI proxy to configure provider models and control Aviator access through virtual keys, guardrails, rate limits, and request logging.

Prerequisites

Prepare the proxy service and provider credentials before connecting it to Aviator.

  • Obtain the AI proxy bundle from the marketplace and deploy it on a customer-managed host.

  • Record the URL where the proxy service runs.

  • Obtain the provider credentials for the models that are configured through the proxy.

For installation, connection, and troubleshooting steps, see Install and configure the AI proxy.

Enable and connect to the proxy

Enable the AI proxy feature and connect the application to the deployed proxy URL.

To enable and connect to the AI proxy:

  1. From the Administration menu, select New Features > Feature Toggles and enable the AI proxy toggle.

  2. Open System Admin > Proxy Configuration.

  3. Enter the proxy service URL in the Proxy URL field and select Connect.

Configure models, virtual keys, and MCP servers

Configure provider models and virtual keys before selecting them for the proxy connection.

To configure models, virtual keys, and MCP servers:

  1. Select Add Model, then specify the Provider, Model Name, and API Key.

  2. Select Create New Key and assign one or more configured models to the virtual key.

  3. In Proxy Configuration, select the Virtual Key and the model or models to use.

  4. Add each MCP server by specifying its Name, Alias, Source URL, and Transport Type.

Configure guardrails, rate limits, and logging

The proxy configuration provides controls for request and response inspection, request rates, and activity review.

  • Guardrails can block input requests or output responses that match configured sensitive terms or regular expressions. An optional AI judge can evaluate content against a configured minimum score.

  • Rate limits can apply to individual users or user groups, with per-minute or per-hour thresholds. The limits count backend calls rather than prompts.

  • Logs display request and response data and token usage.

See also