Roboflow inference inside Codex

Roboflow’s October 19 guide demonstrates inference: applying a trained model to an image. Its inspection example starts by listing available trained models and choosing an exact model identifier. The user supplies an original image, requests predictions, then asks for those same predictions to be drawn onto the image. This exposes the locations behind the result table. The illustrated task is finding defects in ceramic tiles.

The official repository documents a CLI installation route: register its marketplace, restart Codex, open the plugin browser, select the Roboflow source and install the plugin. Enable it if it remains disabled. The bundled connection uses a Roboflow sign-in flow, so it needs no API key. A client unable to complete that flow needs the separately configured key-based alternative.

The repository separates two responsibilities. Its connected server provides live, authenticated access to Roboflow tools and APIs. The bundled skills supply product guidance and workflow instructions. Merely setting an API-key environment variable does not configure authentication for the bundled server connection.

The guide recommends reviewing predictions and evaluating representative labeled images before operational use. Confidence scores do not establish accuracy.