Skip to main content
Every template and agent names one model. It comes from one of two sources: Either way the harness reaches models only through paddock’s model endpoint, which routes each request by the model name it carries. No provider key, paddock’s or yours, is ever present in a sandbox.

The catalog

Each model runs on one harness: Claude Code runs Anthropic models and Codex runs OpenAI models. Naming a model of another harness in a template or an agent is 400. models.list() returns every model your project can use, platform models first, then each of your providers’ models, in one page:

Bring your own provider

Add a model provider to your project with your own key, and the models you choose from it become project models.

How the key is verified

Creating a provider, changing its key or base URL, and both test operations verify the key first, and a create or update whose verification fails changes nothing:
  1. List the provider’s models with the key (GET /v1/models for Anthropic, GET /models for OpenAI). This spends no tokens.
  2. If a compatible endpoint lists no models, send one minimal request to testModel.
Test a key before you store it, or test a stored provider again:
A test answers { ok, method, models, error }: method is list or request, models the IDs the provider listed, and error what the provider answered when ok is false. Testing a stored provider also records its status.

Use a project model

Name it provider/model wherever a model is named: a template’s model and small_model, or an agent’s model settings.
  • The name is split at the first /: the provider, then the model ID sent to the provider as it is.
  • A name with / that is not a model of one of your providers, or a model of another harness, is 400.
  • A model change reaches running sessions between turns; a turn always finishes on the model it started with.
  • A new key or base URL applies from the next request.

Failures

There is no fallback: a request for a project model goes only to that provider, with your key. Fix the key with modelProviders.update(name, { apiKey }); the next successful request or test marks the provider ok.

Manage providers

  • update changes only the fields you send; models replaces the whole list.
  • Removing a model, or deleting a provider, that a template version or a live agent names as its model or small model is 409 conflict (model_in_use).
The console’s Models page does the same, and its model pickers offer platform and project models together; see Console.