Providers

For the providers: config block itself — the fields each provider accepts, defaults, and endpoints — see Configuration. This page covers how a model string is routed to a provider, and the specifics of running against Azure OpenAI and OpenAI.

Model Routing

The prefix in a model string determines which provider handles the call:

Model string	Provider	Notes
anthropic/claude-sonnet-4-6	Anthropic	Native SDK, extended thinking supported
openrouter/deepseek/deepseek-chat	OpenRouter	OpenAI-compat, thinking not supported
ollama/qwen3:8b	Local Ollama	OpenAI-compat via /v1/chat/completions. Note the model-string prefix is ollama/, not ollama-local/ — the ollama-local name is only the providers: config key (see Configuration), chosen to read clearly next to ollama-cloud. Using ollama-local/ in a model string doesn’t match any known prefix and fails as an unrecognized provider.
ollama-cloud/gemma3:27b	Ollama Cloud	Native Ollama /api/chat
google/gemini-2.0-flash	Google Gemini	OpenAI-compat
azure/gpt-4o	Azure OpenAI	OpenAI-compat; gpt-4o is the deployment name
openai/gpt-5	OpenAI	Native OpenAI API, OpenAI-compat, thinking supported on reasoning-capable models
yolo/some-model	Yolo (custom endpoint)	OpenAI-compat, base_url from providers.yolo
claude-sonnet-4-6	Anthropic	Bare name (no prefix) defaults to Anthropic

Azure OpenAI

For Azure, the model string suffix is the deployment name you set up in Azure AI Foundry (not the underlying model family name). If you deployed GPT-4o and named the deployment gpt-4o, the model string is azure/gpt-4o. Different deployments of the same underlying model can have different names.

agent.yaml

name: my-azure-agent
model: azure/gpt-4o           # deployment name from Azure AI Foundry
max_tokens: 4096
model_fallbacks:
  - azure/gpt-4o-mini         # cheaper fallback deployment
  - anthropic/claude-haiku-4-5-20251001  # cross-provider fallback

Azure’s API is OpenAI-compatible. The differences handled internally are the endpoint URL format, the api-key request header (instead of Authorization: Bearer), and the max_completion_tokens parameter (Azure’s chat completions API, like OpenAI’s, rejects max_tokens for reasoning-family deployments — the provider sends max_completion_tokens on the wire regardless of deployment, translated transparently from the agent’s max_tokens field). Extended thinking is not available on Azure OpenAI.

The key name in providers: config must match the prefix in the model string exactly.

OpenAI

openai/ reaches OpenAI’s own API directly — as opposed to routing through yolo/ with a custom base_url, or through azure/. The model string suffix is OpenAI’s own model name (openai/gpt-5, openai/gpt-4o), not a deployment name.

agent.yaml

name: my-openai-agent
model: openai/gpt-5
max_tokens: 4096
model_fallbacks:
  - openai/gpt-4o-mini                    # cheaper fallback, same provider
  - anthropic/claude-haiku-4-5-20251001   # cross-provider fallback

Defaults to base_url: https://api.openai.com/v1; override it in providers.openai.base_url to point at a proxy or OpenAI-compatible endpoint. Auth is Authorization: Bearer. Like Azure, the OpenAI provider sends max_completion_tokens on the wire unconditionally (translated from the agent’s max_tokens field) since reasoning-family models reject the classic max_tokens parameter outright.

Thinking is supported, on reasoning-capable models. thinking_level maps to OpenAI’s reasoning_effort on gpt-5, gpt-5.6, gpt-6-astra (and later), and the o1/o3/o4 family — see Reasoning effort for the exact value mapping and the xhigh clamp on older models. Classic chat models (gpt-4o and similar) have no reasoning-effort parameter at all and silently ignore thinking_level, same as every non-Anthropic, non-reasoning-capable provider; the gateway logs a load-time warning if an agent’s model can’t use the thinking_level it’s been given.

Reasoning-family models and token budgets

Reasoning-family models/deployments (gpt-5 and later, o1/o3/o4) — on both azure/ and openai/ — spend part of their max_tokens budget on hidden internal reasoning before producing any visible output. A budget that’s fine for gpt-4o (e.g. max_tokens: 200) can come back with stop_reason: length and zero visible text on gpt-5 because reasoning consumed the whole budget. Set max_tokens generously (2000+) for reasoning-family models/deployments to leave headroom for actual output.