Model Providers

Supported ProvidersCopy Anchor Link

Agents and judges send every model call to the endpoint in [model]. Nothing is configured by default.

Picking a ProviderCopy Anchor Link

[model]
base_url = "http://localhost:4000"
api_key  = "secret:litellm_api_key"

[model.roles]
agent = "claude-opus-5-5"
judge = "claude-sonnet-5-5"

LITELLM_BASE_URL and LITELLM_API_KEY override the file, and --model on agent-env run changes the agent's model for one run.

LiteLLMCopy Anchor Link

pip install 'litellm[proxy]'
litellm --config litellm.yaml   # serves on port 4000

A LiteLLM proxy serves any model through one endpoint that speaks both the OpenAI and Anthropic APIs, so every agent works through it. List your models in litellm.yaml as its guide shows. Agents in a remote sandbox, such as Modal's, need a base_url they can reach, not localhost.

Hugging FaceCopy Anchor Link

[model]
base_url = "https://router.huggingface.co/v1"
api_key  = "env:HF_TOKEN"

[model.roles]
agent = "zai-org/GLM-5.3"
judge = "openai/zai-org/GLM-5.3"   # a direct judge, use_agent_judge: false

The token needs the Make calls to Inference Providers permission. The router speaks only the OpenAI API, so an agent built on the Anthropic SDK needs a LiteLLM proxy in front of it. Inference Providers covers choosing models and billing.