Providers

A provider is an adapter that connects ra to a language model. It translates ra's internal message format into the API calls that each model vendor expects, and normalizes the streaming responses back into a common format.

ra ships with providers for:

Provider	Models	Config key
Anthropic	Claude 4.5, Claude 4, Sonnet, Haiku	anthropic
OpenAI	GPT-4.1, o3, o4-mini	openai
Google	Gemini 2.5 Pro, Flash	google
AWS Bedrock	Claude via AWS	bedrock
Azure OpenAI	GPT-4.1 via Azure	azure
Ollama	Llama, Mistral, local models	ollama

Switching providers

Change the provider with a flag:

ra --provider openai --model gpt-4.1 "Explain this error"

Or in config:

agent:
  provider: openai
  model: gpt-4.1

The agent loop, tools, middleware, and skills all work the same regardless of which provider you pick. Only the model changes.

How providers work

Every provider implements two methods:

• stream() — returns an async iterable of chunks (text, thinking, tool_call_start, tool_call_delta, tool_call_end, done)
• chat() — returns a complete response (used internally by some features like compaction summarization)

The streaming protocol ensures consistent behavior across vendors. Every stream ends with a { type: 'done' } chunk that carries token usage data.

Credentials

Each provider needs its own API key. Set them as environment variables or in your config:

app:
  providers:
    anthropic:
      apiKey: ${ANTHROPIC_API_KEY}
    openai:
      apiKey: ${OPENAI_API_KEY}

Environment variables are interpolated at load time. See Configuration for the full syntax.

Extended thinking

Some providers (Anthropic, OpenAI) support extended thinking — the model reasons through a problem before responding. ra surfaces thinking tokens in the stream so you can watch the model's reasoning in real time.

agent:
  thinking: medium  # off, low, medium, high, adaptive

See Context Control for more on thinking modes.