Providers

Condukt uses ReqLLM under the hood, which means it speaks to 18+ LLM providers without bespoke client code. You pick a provider by setting the :model option to a provider:model identifier.

Common identifiers

See the ReqLLM docs for the full list and their option semantics.

API keys

ReqLLM auto discovers keys from the environment using the conventional variable names:

export ANTHROPIC_API_KEY="sk-ant-..."
export OPENAI_API_KEY="sk-..."
export GROQ_API_KEY="..."

You can also pass :api_key per agent or set it under config :condukt.

Overriding the base URL

:base_url lets you point at a self hosted gateway, an OpenAI compatible server, or a local LLM:

MyApp.LocalAgent.start_link(
  model: "openai:llama-3-70b",
  base_url: "http://localhost:11434/v1"
)

Local models with Ollama

Condukt ships with Condukt.Providers.Ollama for running prompts against a local Ollama server. This is convenient for development, evaluation, or offline use:

MyApp.LocalAgent.start_link(
  model: "ollama:llama3.2",
  base_url: "http://localhost:11434"
)

Choosing a model

A few rules of thumb:

• For production coding agents, the latest Anthropic Sonnet or OpenAI reasoning models tend to give the best tool use behaviour.
• For local development and tests, smaller open weights models on Ollama are fast and free.
• Lower the :thinking_level (:low, :minimal, :off) on cheap models that do not benefit from extended reasoning.

Switching mid project

Because the model is a runtime option, you can swap it without touching the agent module. This makes A/B comparisons easy:

for model <- ["anthropic:claude-sonnet-4-20250514", "openai:gpt-4o"] do
  {:ok, agent} = MyApp.CodingAgent.start_link(model: model)
  {:ok, response} = Condukt.run(agent, "Refactor this function...")
  IO.puts("#{model}\n#{response}\n")
end