How OpsCodex uses AI

Every AI output — a drafted model, a review suggestion, generated documentation — is a proposal. Nothing moves through the approval → review → publish workflow, and nothing is published, without an authenticated human signing off. There is no fully-automated path from a model's output to an approved process.

OpsCodex does not bundle a model. An administrator configures the provider and API key — Anthropic, OpenAI or Google, or a private/self-hosted endpoint. Requests are made with your own key against your own provider account, so the provider relationship and its data terms are yours. In an on-premise deployment the request goes directly from your instance to the provider you chose; it does not pass through any service we operate.

OpsCodex trains and fine-tunes no models and adds no learning loop of its own. Whether the provider you configure may use API inputs is governed by your own contract with that provider, not by us. On the major providers' API and enterprise tiers, training on API inputs is off; consumer and free tiers can differ, which is one reason OpsCodex expects an organisational API key rather than a personal account.

By default OpsCodex keeps AI telemetry only: which provider and model ran, token counts, status and timing, and a one-way hash of the prompt used for change-detection. The text of your prompts and the model's responses is not stored — not in the database, not in application logs. An administrator can switch on two things: the chat history in the editor, so the reasoning behind an AI-made change stays on record, and the questions asked in the portal, for usage analysis. Even then the content is stored only in your own database, inside your environment, and never reaches us. Switching the chat history off again deletes the stored conversations. AI telemetry is tied to its process draft and is deleted together with it.

AI is inert until an administrator configures a provider key — the deterministic governance product works fully without it. AI-assisted features are opt-in, and you configure exactly the one provider you permit; no other is reachable.

Because the provider endpoint is configurable, you can point OpsCodex at a private, regional or self-hosted model — any endpoint with an OpenAI-compatible interface, such as an internal gateway or your own hosted model — so that prompts never leave your perimeter.

Each AI invocation records a telemetry entry — when AI ran and how much — giving you a reviewable trail without exposing any content.

Questions about AI data handling? Email [email protected]. See also our Security & trust page.