You may already have a preferred model provider, or want an AI gateway for more choice. Before connecting it, understand which service provides each capability, where permissions apply and who charges for usage.
Separate three roles
The client is where you handle conversations and tasks. Your official account supplies capabilities within its entitlements. A third-party provider supplies model inference through an API. These roles can work together without sharing the same entitlements.
Codex Pass helps connect your chosen models and manage their configuration within your existing way of working.
Prepare the connection information
Have your provider’s API address, API key and available model identifiers ready. If you use a gateway, follow its connection information. Model names, protocols and supported capabilities need to match the service.
Start with one model you need and a simple task. Once that works, add models as your work requires. A smaller initial setup makes unexpected behavior easier to isolate.
Account features and API features remain distinct
Original capabilities such as voice and images remain subject to your official account and the client’s support. A custom API key does not turn a third-party model into an official subscription or grant its benefits.
Supported inputs and tools depend on the model and service. Choose according to the task, rather than the name alone.
Keep costs distinct too
The provider you choose supplies third-party inference and sets its billing rules. Codex Pass plans are separate from the API charges of your model provider.
Logs, session and project cost records help you review usage. Compare task outcomes alongside request records to decide whether a model fits your work.
Keep clues when a connection fails
Check the address, model identifier and credential configuration, then inspect the relevant request records. Avoid sharing screenshots or logs containing a full API key.
For issues involving the local environment or tool connections, Doctor checks configuration, connectivity and runtime state. Preserving the context around a failure helps narrow the investigation.