Models

model_settings is the block where you configure the language model and provider-specific fields for the parts of the workflow configuration that contain it. The orchestrator and service_agents use this block as described in the Agents section. The same fields also appear under message_processing.compacting and tool_display. See Message Processing and Tool Display Templates for more details.

Example

Minimal orchestrator.model_settings with openai:

{
  "name": "Example Workflow",
  "orchestrator": {
    "name": "Example Workflow",
    "model_settings": {
      "ai_provider": "openai",
      "name": "gpt-5-mini",
      "context_size": 400000,
      "api_key": {
        "owner_type": "project",
        "owner_id": "<your-project-id>",
        "path": "secret/Custom/<openai-api-key-secret-id>/value"
      }
    },
    "system_prompt": "You are a helpful assistant."
  }
}

Provider choice

ai_provider	Typical use
openai	OpenAI or OpenAI-compatible HTTP APIs
azure	Azure OpenAI
anthropic	Claude models
chariot	Models hosted and served through Chariot

Provider-specific settings

Use exactly one of the following shapes per model_settings object. For external providers (openai, azure, anthropic), supply api_key as a custom Chariot Secret so the key stays managed by Chariot instead of inlined in your configuration.

Chariot Secret format

An api_key field accepts a custom Chariot Secret reference identifying a secret stored in your organization, project, or user scope. The owner_type must be one of "organization", "project", or "user".

"api_key": {
  "owner_type": "project",
  "owner_id": "<your-project-id>",
  "path": "secret/Custom/<secret-id>/value"
}

OpenAI openai

name is required. This example also sets the optional reasoning_effort.

"model_settings": {
  "ai_provider": "openai",
  "name": "gpt-5-mini",
  "reasoning_effort": "medium",
  "api_key": {
    "owner_type": "project",
    "owner_id": "<your-project-id>",
    "path": "secret/Custom/<openai-api-key-secret-id>/value"
  }
}

Azure OpenAI azure

"model_settings": {
  "ai_provider": "azure",
  "name": "gpt-5-mini",
  "azure_endpoint": "https://your-resource.openai.azure.com",
  "api_version": "2024-02-15-preview",
  "api_key": {
    "owner_type": "project",
    "owner_id": "<your-project-id>",
    "path": "secret/Custom/<azure-api-key-secret-id>/value"
  }
}

Anthropic anthropic

"model_settings": {
  "ai_provider": "anthropic",
  "name": "claude-sonnet-4-5",
  "api_key": {
    "owner_type": "project",
    "owner_id": "<your-project-id>",
    "path": "secret/Custom/<anthropic-api-key-secret-id>/value"
  }
}

Chariot-hosted models

Two patterns are supported for calling a model that's served by your Chariot environment.

Option 1: ai_provider: "chariot"

Set name to the model name in your Chariot catalog. project_name and organization_name come from your Chariot project and organization. Alternatively, if your platform team gives you a direct inference endpoint, set model_id and external_base_url instead of the three catalog detail fields.

"model_settings": {
  "ai_provider": "chariot",
  "name": "openai/gpt-oss-120b",
  "project_name": "your-project",
  "organization_name": "your-organization"
}

Two more optional fields are supported on this provider:

• max_api_retries: How many times Chariot retries failed HTTP calls to the inference endpoint when talking to your hosted model.
• openai_reasoning_effort: Turns extra reasoning up or down on Chariot-hosted models that support this control (gpt-oss-120b and gpt-oss-20b).

Option 2: ai_provider: "openai" with chariot_proxy

Use the OpenAI-compatible shape — which lets you set OpenAI-specific fields like reasoning_effort — and route the request to a Chariot-hosted model. Chariot uses your session's bearer token as the API key and reads the inference endpoint from the model record, so api_key must be omitted. Inside chariot_proxy, you can substitute model_id for the three catalog detail fields the same way that was explained in Option 1.

"model_settings": {
  "ai_provider": "openai",
  "name": "openai/gpt-oss-120b",
  "reasoning_effort": "medium",
  "chariot_proxy": {
    "name": "openai/gpt-oss-120b",
    "project_name": "your-project",
    "organization_name": "your-organization"
  }
}

Additional fields

context_size (optional)

Context window size in tokens. Message Processing uses it with context_threshold when compacting is enabled. If this field is omitted, Chariot automatically determines the context window size for some common models.