Models and connections

Choose how your agent reasons, connect the provider, and understand credentials, policy, and readiness.

A model profile tells Core which model to call and how. It includes the provider connection, model identity, reasoning settings, limits, and declared capabilities. An agent selects a profile; it does not carry a raw provider key in its instructions.

Register a model profile

An owner or administrator can open Connections and register a model provider profile. Choose the service and review its endpoint and wire protocol. Enter the exact model ID supplied by that provider; Core does not maintain a model catalogue in this form.

Field	Purpose
Service and wire protocol	Choose the provider and the request format it accepts.
Endpoint URL	The destination the connection will call. Use the endpoint required by that service.
Region	Required for services whose requests are region-specific.
Authentication	Choose the supported authentication method. Enter a new credential or reuse an existing connection binding.
Model ID	The provider's requested model identifier.
Resolution and pre-resolved model	Control how the requested identifier is resolved; use the provider's supported identifiers.
Reasoning	The provider default or a supported explicit reasoning setting.
Context window and max output	The model's prompt-plus-response capacity and maximum response size. Max output must be smaller than the context window.
Tools, structured output, and streaming	Declare what the selected model actually supports. These affect which agent features are valid.

Copy token capacities from the selected model's documentation instead of guessing. The form accepts compact values such as 200K. Declaring a capability does not make an unsupported model implement it.

Understand readiness

There are several independent checks. Registration creates the profile. Connection approval confirms the destination and connection setup. Execution policy permits use in this environment. Use grants permit the person or key starting work to use the profile and its dependencies.

A profile can be visible before all of those checks pass. The agent's publication dialog distinguishes blockers to publication from blockers to starting a run. Open the profile in Connections and the agent's Access section to see the reason, rather than repeatedly republishing it.

Bind it to an agent

Choose the profile when creating an agent, or change it in the agent's Model section. Save the draft and publish a new version. Inspect Model details to confirm the exact provider/model combination and limits.

A tool-using agent needs a profile that supports tools. Streaming requires streaming support and a written-answer output. A structured result needs a compatible model configuration. Test the resulting published version with representative inputs.

Secrets and configuration

Connections also groups integration secrets and configuration. A secret is a protected credential used by a connection or tool. Configuration supplies shared settings. Neither is a knowledge collection, and neither should be pasted into a prompt as a substitute for a proper binding.

Use the controls available on the connection to manage its lifecycle. Some configuration operations are provided through the developer interfaces; see Models, secrets and configuration. A displayed preview control is not confirmation that a live change was made.

Changes and failures

Changing a credential, retiring a profile, or changing execution policy can affect whether dependent runs are allowed to call it. Recheck access after changes. A different model can also require different token limits and capability settings.

Advanced retry and fallback routes belong in the SDK or definition API. A fallback is an explicit model route, not a promise that Core can use any available provider after a failure. See Errors and retries for integration behavior.