Tauon

A minimal Python agent framework built on tau. Think of it like flue of tau instead of pi.

Beta software. Tauon is a thin layer over tau-ai, which has not yet published a stable public SDK. Until that SDK is released, internal APIs may change without notice. Not recommended for production use.

Quick start

from tauon import define_agent, define_tool, use_model, use_tool


@define_tool
def get_weather(city: str) -> str:
    """Get the current weather for a city."""
    return f"Sunny and 22°C in {city}."


@define_agent
def WeatherAgent() -> str:
    use_model("gpt-4.1-mini")
    use_tool(get_weather)
    return "You are a weather assistant."

export OPENAI_API_KEY=...
uv run tauon run examples/weather.py --message "What's the weather in Paris?"

Prompts

Long system prompts are better kept in a markdown/text file than returned from the agent body. Load one with use_prompt():

from tauon import define_agent, use_model, use_prompt


@define_agent
def WeatherAgent() -> None:
    use_model("gpt-4.1-mini")
    use_prompt("prompt.md")

Relative paths resolve against the directory of the module that defines the agent, so tauon run works from any launch directory. Agents defined without a module file (REPL, -c) fall back to the current working directory. use_prompt() and returning a prompt string from the agent are mutually exclusive — using both raises a RuntimeError. See examples/markdown_prompt.py.

Models and environment variables

Tauon resolves models from Tau's built-in provider catalog. API keys are read from environment variables only; Tauon does not load saved credentials from ~/.tau/credentials.json or provider preferences from ~/.tau/providers.json.

use_model() patterns

# Bare model name: resolved against the default provider (openai).
use_model("gpt-4.1-mini")  # needs OPENAI_API_KEY

# provider/model syntax: required when the model ID contains "/" or when you
# want to force a specific provider.
use_model("openai-codex/gpt-5.6-luna")  # needs OPENAI_CODEX_ACCESS_TOKEN
use_model("openrouter/qwen/qwen3-14b")  # needs OPENROUTER_API_KEY
use_model("anthropic/claude-sonnet-4-6")  # needs ANTHROPIC_API_KEY

Provider environment variables

Bare model names (e.g. use_model("gpt-4.1-mini")) resolve against the default provider, which is openai. Use the provider/model syntax for every other provider. For OpenRouter the prefix is required because model IDs contain /, e.g. use_model("openrouter/qwen/qwen3-14b").

Model names not yet in Tau's built-in catalog fall back to a plain OpenAI-compatible provider using OPENAI_API_KEY (a warning is logged). This keeps newly-released models usable immediately, but it also means a typo'd model name is sent to OpenAI rather than failing at startup — use the provider/model syntax to force a specific provider.

For a provider not in the catalog, or to override the endpoint explicitly, pass api_key and base_url directly to run_agent() as shown in examples/custom_provider.py.