Setting up

Add Imp to a Mix project:

def deps do
  [
    {:imp, "~> 0.5"}
  ]
end

or, for a script or a Livebook notebook:

Mix.install([{:imp, "~> 0.5"}])

Imp needs Elixir 1.19 and a C++ compiler for one dependency (erlexec).

Connecting to a model

Imp reaches models through ReqLLM. We give Imp.req_llm/2 a "provider:model" string and a key:

lm = Imp.req_llm("openai:gpt-5.4-mini", api_key: System.fetch_env!("OPENAI_API_KEY"))

Any ReqLLM provider string works in its place: "anthropic:...", "google:...", "openrouter:...", a local server, and more. Every example in this guide runs unchanged with any of them, except that the agent in Tools and agents needs a model with native tool calling. We used gpt-5.4-mini for the outputs shown; yours will differ a little, because models do.

lm is a plain struct. We'll pass it to each program with lm:, which keeps the dependency visible in the code and makes it easy to swap. When we'd rather not repeat it, Imp.configure(lm: lm) sets a default for every program, and Imp.context/2 overrides it for one block of code.

Let's check that the model answers, with the smallest program there is:

hello = Imp.predict("question -> answer", lm: lm)

{:ok, prediction} = Imp.call(hello, %{question: "What is the capital of France?"})
Imp.get(prediction, :answer)
#=> "The capital of France is Paris."

With a model connected, we can write the router.

Next: Your first program →