ReqAI

ReqAI is a lightweight Elixir client for LLM APIs, built on Req.

{:ok, %Req.Response{}, body} =
  ReqAI.Provider.OpenAI
  |> ReqAI.Provider.new()
  |> ReqAI.generate(model: "gpt-5.4-mini", input: "Say hello")

%{"output" => [%{"content" => [%{"text" => text}]}]} = body

IO.puts(text)

See the Getting Started guide for configuration and usage examples.

Philosophy

• Preserve provider-native formats. Use the request and response formats defined by each provider’s API so your code maps directly to its documentation and examples.
• Keep the core small. Focus on a few powerful abstractions and a handful of popular providers.
• Make extension easy. Support custom providers, telemetry, and request and response body transformations so applications can bring their own conventions.
• Treat observability as a first-class feature. Include useful telemetry for common events and make it easy to add application-specific instrumentation.

Features

• Generate and stream with one calling convention across providers
• Adapters for Anthropic, OpenAI, Gemini, xAI, and OpenRouter. BYO using the Provider behaviour
• Telemetry for lifecycle, duration, token usage, time to first chunk, etc. BYO using the Telemetry behaviour
• Full control of the underlying Req client: auth, retries, timeouts, testing
• Optional translators for your own request, response, error, and event shapes

Streaming

Use the same provider and request with a callback that receives each decoded event and updates an accumulator:

{:ok, response, events} =
  ReqAI.stream(provider, request, [], fn event, _response, events ->
    {:cont, [event | events]}
  end)

events = Enum.reverse(events)

Return {:halt, acc} to stop consuming the stream. Events retain the provider's structure, with SSE JSON payloads decoded in event.data.

Installation

Add ReqAI to your dependencies in mix.exs:

{:req_ai, ">= 0.0.0"}

Requires Elixir 1.18 or later and Req 0.8.0 or later.