A minimal terminal AI agent written in Rust.

cargo install --path .

Create tiny.json in your project directory or ~/.tiny/config.json:

{
  "provider": "openai",
  "api_key": "sk-...",
  "model": "gpt-4o-mini"
}

If no provider is configured, tiny uses OpenAI when api_key or OPENAI_API_KEY is set. Otherwise it starts with the local llama_cpp provider.

For llama.cpp, start llama-server and point tiny at its OpenAI-compatible chat completions endpoint:

{
  "provider": "llama_cpp",
  "base_url": "http://127.0.0.1:8080",
  "model": "local"
}

Tool calls require a llama.cpp chat template/model that supports OpenAI-style function calling, for example by starting llama-server with --jinja.

For oMLX, start the local server and configure its OpenAI-compatible endpoint. api_key is optional when oMLX auth is disabled; if omitted, tiny reads OMLX_API_KEY from the environment.

{
  "provider": "omlx",
  "base_url": "http://127.0.0.1:8000",
  "api_key": "omlx",
  "model": "gemma-4-26b-a4b-it-4bit"
}

tiny          # start or resume a session

Apache-2.0